2025 IU POP-UP STORE [Found at Eight] Official MD Notice 📍 Operating Period : September 10 (Wed) – September 21 (Sun), 2025 (※ Closed on September 15 (Mon)) 📍 Venue : THE HYUNDAI SEOUL 5F, EPIC SEOUL 🔗 🔗 #아이유# #IU# #IU_POPUP# #Found_at_Eight#

2025 IU POP-UP STORE [Found at Eight] Official MD Notice 📍 Operating Period : September 10 (Wed) – September 21 (Sun), 2025 (※ Closed on September 15 (Mon)) 📍 Venue : THE HYUNDAI SEOUL 5F, EPIC SEOUL 🔗 🔗 #아이유# #IU# #IU_POPUP# #Found_at_Eight#

[ Marginalia ] Two Hundred and Thirty-Seven. Her name in the file is Mei. Not her name, the one they gave her afterward, so she could be spoken about without being found. At four she was behind in language, in writing, in drawing. Her crayon houses had no doors. The doctors put a word on it that fewer than two hundred and thirty-seven people in the world have ever been given. A gene called CHD3, misspelled by one letter in the earliest draft of her brain. Two hundred and thirty-six others, somewhere, in rooms her parents would never enter. That was the number they held on to. A number is a kind of company. Jason kept the medical records in a folder on his desktop. He opened it more often than he needed to. In the WeChat group, forty parents, all of them awake at two in the morning, all of them typing, a name surfaced in 2023 the way a name surfaces in a city: from someone who knew someone. Shanghai. Xinhua Hospital. A laboratory that had made mice with his daughter's exact mistake inside them. He wrote the email at night. He attached everything: the scans, the reports, the sentence saying he was willing and able to bear the costs. Then he sat there, the way you sit after posting a letter you can't take back. Thirteen minutes. He has counted those thirteen minutes many times since. He has divided them, turned them over, held them up to the light. In thirteen minutes a man can boil water. In thirteen minutes a man on the other side of a country can read a child's whole life and decide what to do with it. Later Jason would understand what the speed meant, that the answer had been waiting long before the question. That somewhere a laboratory had been standing at a window for two years, watching the road, waiting for a father to walk up it carrying a folder and a checkbook. They paid. First for the mice. Then for the monkeys. Then one hundred and thirty thousand dollars into a researcher's personal account, the way you'd pay a man to fix a roof. Here is what nobody said out loud in the consent conversation, though everything else was said, carefully, at length: the worst case is that she dies. They spoke of risks the way a hotel speaks of the weather. The dose went in through the spinal fluid, adenovirus, thousands of times a vaccine's load, because you cannot edit one cell when what you need is a million. Twenty-seven years ago in Arizona there was an eighteen-year-old named Jesse Gelsinger, and on the fourth day his organs stopped, one after another, and the whole world was supposed to have learned something. Somebody did learn something. They learned where the reviews are looser. Because there are two doors in that country. One goes through the drug regulators. The other, for research the hospital calls non-commercial, curiosity's own door, goes through a committee down the hall. The committee that approved Mei never read the primate safety report. This is not an oversight. Oversights get corrected. A door that stays open for thirty years is not a door somebody forgot to close. Mei was not dying. That is the part her mother cannot put down. She was six, she was behind in drawing, she was not dying. A year after she died, the paper came out in Nature. Beautiful mice. Beautiful monkeys. A breakthrough. What had been quietly removed: the human data, and the line thanking a family in the acknowledgments. Their daughter had been edited out of the sentence that thanked them for her. So they went and asked for the paper to be pulled. Not for her. For the parents reading it that morning, in forty group chats, at two in the morning, thinking: there's hope now. Two hundred and thirty-seven, minus one. Nobody has updated the number. Nobody was ever going to. — Aric

Twenty five years ago today a 24 year old equities trader tied a red bandana over his face and went back up into a burning tower until it killed him. Welles Crowther worked for Sandler O'Neill on the 104th floor of the South Tower. Nyack, New York. Boston College lacrosse. When he was six his father handed him a red handkerchief and told him to keep it in his back pocket, and he carried one every day after that. Teammates knew him by it. He wore one under his lacrosse helmet. At 16 he joined his father's volunteer fire company in Upper Nyack. By 2001 he had a trader's salary and a trader's future and he had already put in an application to the FDNY, because he wanted out of finance and into a truck. 9:03 a.m. United 175 goes into the South Tower, floors 77 through 85. The 78th floor sky lobby is right in the impact zone, packed with people waiting for express elevators down. Most of them are killed instantly. The ones still alive are lying in the dark in burning jet fuel and collapsed ceiling, and none of them know that one stairwell is still open. Judy Wein, badly injured, hears a voice cutting through it. A young man with a red bandana over his nose and mouth, carrying a fire extinguisher. He puts out flames. He finds the intact stairwell. Then he says, everyone who can stand, stand now. Everyone who can help others, find someone to help. He lifts an injured woman onto his back and carries her down seventeen flights to the 61st floor, where the elevators are still running. Then he turns around and climbs back up to 78. He does it again. Ling Young, burned over most of her body, is led down by the same man in the same bandana, with another injured woman in his arms. He hands them off at 61, points them toward the elevators, and goes back up a third time. Somewhere in all of that he called his mother's voicemail. "Mom, this is Welles. I want you to know that I'm okay." The South Tower fell at 9:59. In March 2002 they found his body in the lobby, intact, lying with FDNY firefighters and the jaws of life. He had gotten all the way down and stayed to work the fire department command post instead of walking out the door. His family knew none of it. Then that May his mother read survivor accounts in the paper describing a man in a red bandana who came out of the smoke on the 78th floor and told them to stand up. She recognized her son from a piece of cloth. At least a dozen people are alive because of him, and the number is usually put at eighteen. In 2006 the FDNY made him an honorary firefighter, something almost no civilian ever receives. He was 24 years old.

Her older half-brother died in a car accident. For years, her father wore dark sunglasses every single day just to hide his deep sadness and tears from the world. Shakira was only two. Her father, William Mebarak Chadid, was devastated. The son he lost was William, his eldest child from his first marriage. He was just 19. In 1979, his motorcycle was struck by a car on a Colombian road. The loss changed everything. Her father fell into a deep, lasting depression. Her mother, Nidia del Carmen Ripoll Torrado, became the family’s strength. She supported her grieving husband while encouraging Shakira’s growing talent. Shakira grew up in Barranquilla, Colombia. Her Lebanese father was a jewelry store owner, writer, and journalist. Her mother brought Spanish and Italian heritage into their home. Creativity came early. At four, Shakira wrote her first poem, “La rosa de cristal.” Then her Lebanese grandfather took her to a Middle Eastern restaurant. She heard the traditional doumbek drum. And discovered belly dancing. But her voice wasn’t immediately celebrated. At Catholic school, a choir teacher rejected her. He said her powerful vibrato sounded “like a goat.” Shakira was incredibly shy. Belly dancing became her escape. Every Friday at school, she danced. Slowly, she found the courage to express herself. But she never forgot her father’s sadness. At eight, she turned those feelings into a song. “Tus Gafas Oscuras.” “Your Dark Glasses.” She wrote it for her father. A little girl trying to comfort the man she loved. That empathy followed her into adulthood. At 18, moved by poverty, she started Fundación Pies Descalzos. Its mission was to build schools and provide meals for vulnerable Colombian children. At 26, she became UNICEF’s youngest Goodwill Ambassador at that time. Fame didn’t stop her curiosity either. In 2007, she secretly attended a UCLA history course. She registered as Isabel Mebarak. Cap. Sweatpants. Back of the classroom. During the 2020 lockdown, she earned a certificate in Ancient Philosophy from the University of Pennsylvania. She kept learning. She kept helping. And the little girl who once turned her father’s pain into a song eventually turned her voice toward millions of others. Before Shakira became a global star, she had already learned something bigger than fame— how to turn pain into compassion.

A Haitian-born formerly enslaved woman spent thirty years walking into New York slave markets with cash in her pocket, paying full price, and walking back out the door with signed manumission papers. Her name was Juliette Noel Toussaint. She was born into slavery in the French colony of Saint-Domingue — modern Haiti — around 1786. Her enslavers, the Noel family, fled the Haitian Revolution during the early 1790s and dragged their household slaves with them, first to Baltimore, then to New York. She was a housemaid, a cook, and a nanny before she was ten years old. She knew, better than almost anyone alive in the city of New York in the 1820s, exactly what it meant to be priced, appraised, and legally sold. Then, in 1811, she married a hairdresser. Pierre Toussaint had been born into slavery in Saint-Domingue in 1766. Brought to New York in 1787 by the Bérard family. Apprenticed as a hairdresser as a young man. By his mid-thirties he was the most sought-after hairdresser in the city, styling the hair of the wives of Alexander Hamilton, Philip Schuyler, and John Church at their Upper Manhattan estates. When his enslaver Madame Bérard died in 1807, he was freed. The first thing he did was buy Juliette's freedom out of the Noel household. They married at St. Peter's Catholic Church on August 5, 1811. They moved into a modest two-story wooden house on Franklin Street in lower Manhattan. Pierre worked at the estates six mornings a week. Juliette managed the household accounts. She decided where the money went. New York State had passed the Gradual Emancipation Act in 1799. It was a legal twilight zone. Children born to enslaved mothers after July 4, 1799 had to serve extended terms — twenty-eight years for men, twenty-five for women — before they were free. Enslavers who did not want to wait sold their people to Southern buyers who came up to the harbor docks, loaded them into ships, and hauled them to the cotton and sugar plantations of Louisiana and Mississippi. The market was legal. The paperwork was standard. The clock was running out. But the clock was still running. Juliette watched. She started intervening around 1815. She found people about to be sold South. She found families about to be broken up by a master's unpaid debts. She found Haitian refugees living on the docks near the East River, terrified of being seized by slave-catchers and shipped down the coast. The first time she walked into a transaction, she brought cash. She did not bargain. She did not plead. She placed the money on the table. The seller counted it. The seller signed the bill of sale. The moment the ink dried, Juliette signed a second document. A manumission paper. The property was now a person. She filed the manumission paper at the clerk's office at City Hall the same afternoon. She paid the small processing fee. She kept the receipt in a small wooden box in the back kitchen of the house on Franklin Street. She did this for the next thirty years. She kept no public ledger of the names of the people she bought. She wrote down no biographies. The less written evidence that existed, the safer they were from Southern slave-catchers who came north on legal warrants trying to reclaim "escaped property." The receipts were her only record. When the cholera epidemic swept through the lower wards in the summer of 1832, killing thirty-five hundred New Yorkers and leaving thousands of children orphaned, she pulled her savings again. She helped fund the first Catholic orphanage in the city — the St. Vincent de Paul asylum. Pierre opened their own home on Franklin Street to Black orphans nobody else would take. She paid for pews at the first St. Patrick's Cathedral on Mott Street. When she tried to sit in them one Sunday morning, the white ushers asked her to move to the back of the sanctuary. She moved. She kept funding the parish. She and Pierre took no vacations. They bought no larger house. They purchased no fine clothing. Pierre made enough money styling the hair of Manhattan's oldest aristocratic families to fund a small country estate in Westchester. Juliette decided instead to spend essentially all of it on other people's freedom. She died at the Franklin Street house on May 14, 1851, at approximately sixty-five years old. Pierre died in the same house two years and one month later, on June 30, 1853, at eighty-seven. The St. Vincent de Paul asylum they helped fund moved twice and merged into other institutions across the following century. The house on Franklin Street was torn down. The blocks of lower Manhattan where Juliette walked into slave markets with cash in her purse are now paved over by the skyscrapers of the Financial District. The New York Public Library holds Pierre Toussaint's papers today at the Stephen A. Schwarzman Building at Fifth Avenue and 42nd Street. Roughly twelve hundred documents dated 1793 to 1853. Hundreds of letters from people asking for money, for shelter, for sanctuary. Manumission receipts filed at City Hall between 1815 and 1849. The archive is cataloged under the title Pierre Toussaint papers. The husband's name only. The manumission receipts are still inside. Juliette Noel Toussaint: the woman who bought freedom in secret. If her story stayed with you, drop one word in the comments — Juliette, Franklin, freedom, anything that comes to mind. Tap the like button so more people find this story. The page is small. Every reaction helps us keep telling stories like this one.

She graduated from Princeton in 1992, and one of her professors — a woman who would go on to win the Nobel Prize in Literature — said she was one of the finest creative writing students she had ever taught. Her professor's name was Toni Morrison. And the young woman she praised was about to disappear, almost entirely, into someone else's story... Her name then was MacKenzie Tuttle, and the world would come to know her as MacKenzie Scott. What almost no one realized, watching this poised young Princeton graduate, was how close she had come to never finishing at all. Her family had filed for bankruptcy during her final year of boarding school. At Princeton, on scholarship, she struggled constantly to pay her way, waiting tables to cover her costs. In her sophomore year, broke and in tears and on the verge of dropping out, she was saved by a roommate who found her crying and pressed a thousand dollars into her hands. And once, when she couldn't afford to fix a broken tooth and had been holding it together with denture glue, a local dentist did the work for free, asking nothing in return. She never forgot either of them. Remember that — it matters later. After graduation, with her gift for words and a good head for numbers, she took a practical job at a hedge fund in New York to earn a living. And there she met a man named Jeff Bezos, who had a strange idea about selling books over the internet. She didn't dismiss it. They married in 1993. And in 1994, they packed their belongings into a car and drove across the country to the Seattle area, chasing that idea. They rented a house with a garage. They set up computers. There was no big budget, no team, no guarantee that a single stranger would ever buy a single book. MacKenzie did the accounting. She wrote up business plans and contracts. She answered customer phone calls. She packed boxes for shipping. The company was called Amazon. Its first full year of sales came to about five hundred and eleven thousand dollars. And she was there for every bit of it, from the empty garage onward. But as Amazon exploded into a giant, MacKenzie stepped back from it. She and Jeff had four children, and she turned her focus to raising them — and to the dream she'd held since Toni Morrison's classroom, of being a writer. In 2005 she published her first novel, and it won an American Book Award. Morrison herself praised it as "a rarity: a sophisticated novel that breaks and swells the heart." MacKenzie published a second novel in 2013. She taught writing. She lived, by choice, out of the spotlight. And so, for the better part of two decades, the story of Amazon was told without her in it. Jeff Bezos became one of the most famous men alive. The garage became the stuff of business legend. And in that legend, MacKenzie appeared only rarely and briefly — a wife mentioned in passing, tucked into the corner of someone else's triumph. She never once stepped forward to correct it, or to claim more of the credit than she was given. She let it be. Then, in 2019, everything changed. That January, she and Jeff Bezos announced their divorce. When the settlement was finalized, MacKenzie walked away with roughly four percent of Amazon's stock — a fortune worth somewhere around thirty-six to thirty-eight billion dollars. Overnight, at forty-eight years old, she became one of the richest women on the planet. And what she chose to do next is the real reason you should know her name. Within months, she signed the Giving Pledge, publicly promising to give away the majority of her enormous wealth within her own lifetime. But it was the way she gave it that stunned the entire world of philanthropy. There were no lengthy application forms. No demands for progress reports. No strings attached, no conditions, no control. No grand foundation bearing her name, no press releases trumpeting her generosity. Instead, she assembled a small, low-profile team whose job was to seek out organizations doing serious, unglamorous good in overlooked communities — and then call them up and give them money. The phone calls became the stuff of legend in their own right. So many recipients assumed it had to be a scam. Some asked the caller to repeat the dollar figure, certain they'd misheard. Others just broke down and wept. She gave four hundred and thirty-six million dollars to Habitat for Humanity. She handed the largest single gifts in the entire histories of several Historically Black Colleges and Universities. She sent hundreds of millions to food banks, to rural communities, to tribal colleges, to women's health programs — the kinds of quiet, difficult, deeply human causes that flashier donors so often pass over as too small or too complicated to bother with. In 2025 alone, she gave away more than seven billion dollars, to hundreds of organizations. By many public accounts, that single year of giving outstripped the entire lifetime charitable giving of her famous former husband. Her total, since 2019, now stands at an astonishing twenty-six point three billion dollars, spread across more than two thousand seven hundred organizations. No buildings carry her name. There is no branded foundation. There are no galas thrown in her honor. Today, Forbes ranks her as the third most generous living philanthropist in the entire world — behind only Warren Buffett and Bill Gates. She climbed to that place in a mere six years. She packed the first boxes. She ran the first numbers. She answered the first phone calls. The company that grew out of that little garage is now worth more than two trillion dollars. The garage still stands. Her name was never on it. But what she has done with the fortune that came out of it is now a matter of public record — more than twenty-six billion dollars sent out into the world with no fanfare and no strings, in exactly the spirit of a dentist who once fixed a broken tooth for a struggling student who couldn't pay, and a roommate who dried a young woman's tears with a thousand dollars she didn't have to give. In one of her own essays, MacKenzie Scott described herself as just one small part of a much larger story. The numbers say otherwise. Had you heard MacKenzie Scott's story? Tell us below. And here's one for today: has a stranger's quiet act of kindness ever changed the direction of your life — and did you ever get to pass it on?

here's a prompt to improve your agent harness based on what we've learned at cursor. enjoy # Improve this agent harness's token efficiency You're working on an LLM agent harness: the system prompt, tool definitions, request assembly, context caching, compaction, and retrieval, and how work is split across agents. Make the agent's runs cheaper without making it worse at its job. - Objective: lower price-weighted token cost per completed task. - Constraint: no measurable drop in task quality. Measure per task, not per request. Every turn resends the prefix (tools, instructions, setup, and the conversation so far), so a change that shrinks each request but adds turns can cost more. Weight tokens by billing type: output, uncached input, and cached input are priced very differently. Work in this order: map the harness and measure the baseline, rank the opportunities, make the changes that are safe to make directly, put the rest behind flags or in proposals, then report. Figures below come from one team's production coding agent and its multi-agent experiments. Use them to gauge magnitude, not as targets. One round of these changes (prompt trimming, tool offloading, cache layout, sparse line numbers, subagent tuning) cut that team's overall token cost about 7% with no loss in quality. The larger percentages apply only to the part of the request each change touched. ## Principles 1. Change what the harness sends, not how hard the model tries. Don't ask the model to conserve tokens. A harness that told its model to "take care to preserve tokens and not be wasteful" found it grew reluctant to take on ambitious tasks and sometimes quit, saying it wasn't supposed to waste tokens. 2. Capable models need definitions, not commands. Lists of "DO NOT", "You must", and "Important", and guards against older models' habits, can usually be replaced with plain descriptions of what each tool does. One team cut about two-thirds of its system prompt this way, and the shorter prompt worked across model families. Instruct only on what the model can't know (the product, the environment, the user's processes) and on quirks you've seen in transcripts. 3. Static context is for what most turns need. Everything else should be discoverable when needed. Less up-front context also means less confusing or contradictory information. 4. Expect removals to win. Guardrails written for weaker models, coordination steps that became bottlenecks, and prompting for behavior the model now does on its own all cost tokens. 5. Real usage decides. Evals are a fast proxy, but they skew toward hard problems and miss the real mix of requests. ## 1. Map the harness and measure the baseline Find: - Where requests are assembled, the system prompt, and tool schemas. If a framework or SDK builds requests, find its hooks for message order, cache control, and tool loading. - How tool results are formatted, and how history is kept, trimmed, or summarized. - How subagents or parallel agents are spawned, if any. - Which models and provider APIs are used. From the provider's docs, get the prompt caching behavior (automatic or explicit breakpoints, TTL, minimum cacheable length) and the prices for output, uncached input, and cached input. - Existing logging, token accounting, and evals. If the harness doesn't record per-request token usage by billing type and cache hits, add that first. Everything later depends on it. Then render a few real requests (from logs, or by running representative tasks) and count tokens per section with the model's tokenizer or the API's usage fields. Produce: - Cost share by source × billing type. Sources: system prompt, tool definitions, skill/rule/integration descriptions, user messages, file reads, search results, command and other tool output, history, summaries, subagents. - Static tokens per request, cache hit rate, and turns per task. - Per tool: the share of runs that call it at least once, and its error rate. Read the rendered requests, not just the templates. Duplication, leaked volatile values, and misordered blocks only show up there. Rank opportunities by share of spend × fraction removable ÷ quality risk. ## 2. System prompt and injected context Label every instruction: - Keep: product or environment knowledge the model can't infer, fixes for quirks seen in this model's transcripts, and rules a mode depends on. - Rewrite: commands and emphasis into plain descriptions. Reminders into constraints: "No TODOs, no partial implementations" works better than "remember to finish implementations." Vague quantities into ranges: "generate 20–100 tasks" gets far more ambitious behavior than "generate many tasks." - Delete: things capable models do by default, guards against behavior you haven't seen from this model, text that repeats tool descriptions, and lines that could contradict a user request. Models trained to rank system instructions above user messages will side with the system prompt. - Move: anything per-user or per-request (date, environment, repo state, lists of skills or subagents, user rules) into a user-role setup message after the cache boundary. Audit other injected context the same way. As models improved, the team behind these figures dropped directory trees, pre-retrieved snippets, compressed copies of attached files, lint errors injected after every edit, forced expansion of short file reads, and caps on tool calls per turn. They kept small, high-value facts: OS, repo status, and open or recently viewed files. Skip checklists for open-ended work. The model optimizes the listed items and deprioritizes everything else. ## 3. Tool definitions Tool schemas ride along on every request. Most tools beyond the core set were each needed in under 20% of conversations, and moving them out of static context cut tool-description tokens 60%. Doing the same for integration tools (such as MCP servers), with names in context and full schemas in one folder per server that the agent can search with grep or jq, cut total tokens 46.9% in sessions that used them. - Keep in static context: high-frequency tools (for a coding agent: read, search, edit, shell), tools the model tries to call even when they're absent, and tools a mode depends on. - Offload the rest: leave a name or one-line pointer and make the full schema discoverable on demand. Group related tools so they load together, and put status (such as "needs re-authentication") where the agent will see it. - Tighten what remains: describe behavior and arguments, and drop usage lectures. - Pick the split by testing a few configurations and tracking tokens, cost, latency, tool-call errors, and task success. ## 4. Cache layout Order each request so the reusable prefix is as long as possible: `tool definitions → system instructions → [breakpoint] → setup message (skills, subagents, rules, environment) → [breakpoint] → conversation` - Keep the prefix byte-identical across turns. Use deterministic tool order and serialization, put timestamps and IDs after the boundary, and don't rewrite earlier messages except when compacting. - Use explicit breakpoints if the provider supports them. Otherwise rely on automatic prefix caching with the stable part first. Respect TTL and minimum-length rules. - Switching models mid-conversation throws away the cache (caches are per model and provider) and hands the new model a history it didn't write. When a different model is needed, run it as a subagent with fresh context. Explicit breakpoints plus moving per-request setup after them cut cold cache misses 20%. ## 5. Tool results and other context added during a run - Large outputs (commands, integrations, logs): write them to a file and return the path, size, and a short tail. The agent can tail, grep, or read ranges for more. Truncating loses data, and inlining bloats every later request. Treat long-running terminal sessions the same way. - High-volume formats: look for overhead repeated on every line or item. Numbering every 10th line of a file read instead of every line cut cache-read tokens 1.6% without hurting citation accuracy. Each number costs 3–5 tokens, and agents read tens of thousands of lines per session. Also check repeated absolute paths, verbose JSON keys, ANSI codes, progress bars, and repeated headers. - Good retrieval saves exploration turns. Adding semantic search alongside grep raised codebase question-answering accuracy 12.5% on average and cut the iterations users needed. - Tool errors waste tokens and leave confusing debris in context. Classify expected errors (invalid arguments, unexpected environment, provider error, timeout, user abort), treat unknown errors as harness bugs, and track rates per tool and per model. One focused effort along these lines cut unexpected tool errors 10×. ## 6. Long runs: compaction, subagents, and model mix - Compaction: keep the summarization prompt short and the summary compact, carry forward plan state and remaining tasks, and save the full history to a file the agent can search for details the summary dropped. A model trained to self-summarize from a one-line prompt wrote ~1k-token summaries with half the compaction error of a multi-thousand-token prompt that produced 5k+ token summaries. Untrained models may need more guidance, so test how short you can go. A more expensive summarization model made a negligible difference. - Scratchpads and running notes: rewrite them instead of appending. For repeated work in one environment, a small agent-maintained notes file with a line budget, loaded at start, is a promising way to shorten later runs. - Subagents: fresh context keeps the parent lean, but isolation adds coordination cost (duplicate or stale work). If the model already delegates on its own, remove prompting that pushes it to. Have subagents return short handoffs: what was done, findings, concerns, and deviations. A subagent should use a different model only when the user or harness says so. - Model mix: in large multi-agent runs, workers used at least 69% of tokens, and over 90% in most runs. A frontier planner with cheap workers matched a frontier model doing everything at about one-eighth the cost. Planner choice still changes worker spend. One planner that cost less on its own saw its workers use several times more tokens, and the run cost more overall. Measure the whole tree. - Routing and reasoning effort: send simple turns to a cheaper model or lower effort, and upgrade only when a stronger model is clearly better. A router built this way matched or beat single frontier models on user satisfaction at 41–68% lower cost. - Reasoning continuity: if the API returns reasoning items (including encrypted ones), pass them back on later turns and alert when they go missing. Dropping them cost one reasoning model 30% on a coding benchmark, and it burned tokens reconstructing its plan. ## 7. Fit the harness to each model Adapt to what each model was trained on instead of forcing one shape on all of them. If you've tuned the harness for a similar model, start from that version. - Edit format: use the one the model was trained on (for example, patch-style or search-and-replace). An unfamiliar format costs extra reasoning tokens and causes more mistakes. - Shell or tools: shell-first models fall back to `cat` or inline scripts. Name tools after their shell equivalents (such as `rg`), and if needed add: "If a tool exists for an action, prefer to use the tool instead of shell commands (e.g. read_file over `cat`)." - Literalness: some model families follow instructions literally and others tolerate imprecision. Some spiral on emphasized wording. Strip caps and emphasis for literal models. - Triggers: some models ignore a tool until told when to use it. A literal trigger works: "After substantive edits, use the to check recently edited files for linter errors. If you've introduced any, fix them if you can easily figure out how." - Progress updates: if a model reports progress through reasoning summaries, keep them to 1–2 sentences that note new findings or a change of tactic, and remove instructions about messaging mid-turn. - Quirks worth a targeted line: hedging or refusing as context fills ("context anxiety"), declaring completion early, stopping to ask permission, and calling tools that don't exist. Tie each added instruction to the transcript behavior it fixes. Re-audit when models change, since guidance one version needed can be dead weight for the next. ## 8. Validate - Offline: run a fixed set of realistic tasks before and after, ideally drawn from real usage and phrased the way users actually write (short and ambiguous). Compare task success, tokens, cost per task, turns, and tool errors. Don't ship a change that lowers success. - Online, if you have users: A/B test each change or small bundle. The primary metric is cost per completed task. Guardrails are task success signals, tool-call errors, latency, turns per task, and cache hit rate. For a coding agent, a good success signal is how much agent-written code survives over time. In general, check whether the user's next message moves on or reports a problem. - Ship only when cost drops and no guardrail regresses beyond noise. Record null results. ## What to change directly and what to propose - Change directly, each in its own revertible commit: token and cache telemetry, deterministic serialization and tool order, moving volatile content out of the cached prefix, explicit cache breakpoints, writing large outputs to files instead of truncating, passing back reasoning items that are being dropped, and fixes for recurring tool errors. - Change behind a flag so it can be tested: system prompt edits, tool offloading, output format changes, compaction changes, and subagent prompting. - Propose only: changes to which models run, routing, reasoning-effort defaults, or how work is split across agents. ## Traps - Asking the model to use fewer tokens or do less. - Truncating tool output. - Dropping reasoning items to save input tokens. - Volatile content in the cached prefix, or tool order that changes between requests. - Offloading a tool the model needs on the first turn or tries to call when it's missing. - Emphasis-heavy prompts (MUST, NEVER, IMPORTANT, all caps), especially with literal models. - Forcing a terser output format than the model was trained on. Fewer output tokens can mean less thinking and worse results. - Optimizing raw token counts instead of cost, per request instead of per task, or evals instead of real usage. - Switching models mid-conversation to save money. - Adding coordination layers that become bottlenecks. ## Report back with 1. The harness map and baseline: cost by source × billing type, with the biggest sources called out. 2. A ranked list of changes: layer, what changes, estimated savings and how you estimated them, quality risk, how to validate, and how to roll back. 3. The changes you made, including a system prompt diff with a keep, rewrite, delete, or move reason for each line. 4. A test plan for the flagged changes. 5. Gaps: anything you couldn't find or measure.

Eight Italians, 25 Spaniards, 27 Romanians — and no sign that Revolut paid the ransom Local counts so far: 12 in Ireland, 27 in Romania, 25 in Spain, 8 in Italy, according to the media. Revolut has not published its own country list. It still says only a “limited” number. No evidence it paid a ransom. IAmNotAVillain posted a demand for 6,000 XMR, about $3m, with a 24-hour clock. Revolut told Reuters it had had no direct contact and no demand. The deadline passed. Nothing public followed. Italy is investigating two separate claims. One is settled enough: Revolut did hand over customer files. The other is not: the group’s story that it stole 147 GB from the Interior Ministry. Sole 24 Ore reports that the Postal Police have found no confirmed breach of ministry systems so far. Investigators still do not know whether a real PEC account was taken over or only copied. Giulia Pastorella (Azione) says Undersecretary Ferro used the figure of eight Italian customers to avoid the real question — how a government mailbox was used. She adds that attacks over certified email almost doubled in 2025 (CERT-AgID) and that a black market exists for institutional PEC accounts. How it started. For months, mail arrived from a real Italian government domain. Revolut sent identity documents, selfies, statements and Bitcoin histories. Later, it checked with the agency, blocked the address and told those customers. The app was not broken into. The money did not move. The files did. Three warnings were in that correspondence. The sender was a local-authorities PEC at the Reggio Calabria prefecture. The letter inside was framed as a European Investigation Order, which in Italy only a prosecutor or a judge can issue. Staff then told the sender to put the Cyprus crypto company on the letterhead instead of the Italian branch, and still sent the files. The 680 people on the list are waiting to see whether their documents are offered for sale. The Italian government has not yet said how that official mailbox was used.