🎮 Game agents can demo well but fail in production. 🔍 A 3-day, 7-step AgentLoop method: • Trace evidence, not claims • Evaluate tools & API facts • Fix root causes with Skill guardrails • Regress with datasets & alerts #AIAgent# #GameDev# #AgentLoop# #LLMOps# #CodingAgent#

TL;DR: A new pipeline automatically builds 5,545 RL training tasks for coding agents using only source code itself, no issues or commit history needed, and it prioritizes quality over quantity. Title: CodeMidas: Scaling Agentic Coding RL Environments from Code Itself URL: Key points 🏗️ Auto-builds 5,545 tasks from 3,185 repos across 23 languages and 15 domains 🧪 Generates verifiers via execution-grounded tests, running the reference solution to record expected outputs 🛡️ Three-stage filtering: leakage checks, agent-solution agreement, and rollout difficulty filtering 📈 Big gains after RL training: DeepSWE +11.7pt, ProgramBench +17.0pt, Terminal-Bench +8.5pt 🔍 5k filtered tasks consistently beat 8k unfiltered tasks, proving quality beats quantity 🧠 Trained agents explore more and self-verify more, and these behaviors transfer to external benchmarks I like how simple and practical the core idea is: you don't need issue trackers or dev history to build RL environments, just the code. #CodingAgent# #ReinforcementLearning#

Selecting the right model for a task, and for every subagent it spawns, is a decision you shouldn't have to make by hand on every run. Smart Routing makes it automatic. Turn it on once, and Omnigent routes each task and its subagents to the best model and harness for the job: 🔹 Open Source: an LLM classifier picks the best model 🔹 @databricks: powered by AI Gateway smart routing Learn more 👉 #Omnigent# #OpenSource# #AIAgents# #LLMOps# #CodingAgents#

1 week out: Omnigent & Multi-Agent Coding Meetup in NYC 🗽 Join us at the Databricks NYC office for talks on Omnigent, custom models, and multi-agent orchestration + governance, with speakers from @databricks, @modal, and @cursor_ai. 🙌 📅 Wed, Aug 19 🕒 5:00–8:00 PM ET 🎟️ RSVP: #Omnigent# #AIAgents# #CodingAgents#

2 weeks out ⏳ Omnigent & Multi-Agent Coding Meetup in NYC Join us at the @databricks NYC office with speakers from Databricks, @modal, and @cursor_ai. Expect a primer on Omnigent, a talk on custom models, a panel on multi-agent orchestration and governance, and mingling with bites and dessert. 🙌 📅 Wed, August 19 | 5:00–8:00 PM ET 🎟️ RSVP: #Omnigent# #AIAgents# #CodingAgents#

💰 Same model, same task, but swap the software wrapper (harness) around it and the cost can jump 5x while success rate barely moves. Title: HarnessTax: How Much Does the Harness Matter for Coding Agents? URL: UC Berkeley and Arena compared three coding agent harnesses, Claude Code, Codex CLI, and Pi, across 7 models and 21 model-harness pairs. Here are 3 highlights. 📊 Statistically, harness barely moves success rate Out of 42 within-model harness comparisons, only 1 was statistically significant (chance alone predicts about 2), and none survived correction for multiple comparisons. 💸 Cost can differ 5x at the same success rate On GPT-5.6 Luna, Claude Code cost $0.15/task at 55.6% success, while Pi cost $0.03/task at 53.3% success, nearly identical accuracy, 5x the price. 🪶 The minimal open-source harness holds its own In 9 of 12 model comparisons, a non-vendor harness (Pi or Codex CLI) achieved the top success rate, showing a stripped-down harness can genuinely compete. The practical takeaway: prioritize cost and reliability over flashy add-on features first. #CodingAgents# #LLMCostOptimization#

Turns out the best harness design for a coding agent — planning, tool setup, context management — can flip completely depending on how capable the model is. This paper tested 176 configurations to prove it. Title: An Empirical Study of Harness Design for Coding Agents URL: 🧠 Highlight 1: Context management matters most when resources are scarce At a 32k-token window, managed vs. unmanaged context created a 35.7-point gap in SWE-Bench success rate. Interestingly, the fancy "recall" mechanism was barely ever used and added no accuracy at all. 📋 Highlight 2: Planning's benefit flips with model strength A weaker 30B model gained +11.6 points from adding a planning tool, and its rate of quitting without even attempting an edit dropped from 69% to 28%. A stronger 550B model needed no such help — adding planning there just cut cost by about 30% with no accuracy change. 🔧 Highlight 3: Tool design also depends on the model and task Weaker models need a full dedicated toolset, while stronger models often perform better and cheaper with bash-only access. Even the same model can flip its optimal choice depending on the task type. This really drives home that tuning each harness component to your specific model and task is worth taking seriously. #CodingAgents# #LLM#

🧵 You told the agent on day one to never expose internal DB IDs. Sixty messages later it builds a new auth flow. Does it still remember? Title: Coding Agents Don't Need Longer History — They Need Intent Continuity URL: The article's premise: bigger context windows and RAG don't fix this. What's missing is a mechanism to check whether past intent still applies right now. Highlight ①🔍 A new axis: intent continuity RAG answers "what's relevant?" Longer context solves capacity. Neither answers "is this still true?" The framework splits recovery, validity checking, and application into distinct steps. Highlight ②⚙️ Supersession detection as the verification core When two records share the same target but disagree, the later one wins chronologically — but records in different scopes (prod vs. prototype) both stay active. This verification layer, the author argues, does heavier lifting than retrieval itself. Highlight ③📊 The numbers back it up On 70 synthetic interactions, keyword search hit 0.57 recall and passed only 4 of 8 tasks. The intent-aware approach hit 1.00 recall, zero violations, and passed all 8 — for about 28% more tokens. A quiet but essential lens for running coding agents on long-lived projects. #AIAgents# #CodingAgents#

TL;DR: Same Claude Opus, no harness → $9 outcome in 20 minutes. Same model, full harness → $200 working product in 6 hours. This open-source curriculum teaches why — systematically. learn-harness-engineering "The model is smart. The harness makes it reliable." — 14 lectures and 8 projects covering harness engineering for AI coding agents. ⭐️ 13.7k Key points 🔧 5 subsystems form the harness skeleton · Instructions: AGENTS.md and structured guidance make the repo the single source of truth · State: claude-progress.md + git history enable multi-session continuity · Verification: tests, lint, E2E pipelines prevent false completion claims · Scope: explicit completion criteria enforce one-feature-at-a-time progress · Session Lifecycle: init → implement → verify → commit → handoff as a structured protocol 📚 Progressive learning path From fundamentals (seeing the problem, structuring repos) to advanced patterns (automated loops, graph-based workflows, Human-in-the-Loop). A shared Electron knowledge-base app provides hands-on practice at every stage. ⚡ Instantly applicable to existing projects Copy AGENTS.md, and feature_list.json from the template library into any project root for immediate structured context and scope constraints. The zero-dependency diagnoses harness quality on the spot. 🌐 Current frontier harness designs included Breakdowns of Pi, Claude Code, Codex, and DeepSeek harness architectures (added August 2026), plus loop and graph engineering lectures with templates (July 2026) — the course continuously reflects real-world practice. The shift from "prompt the agent better" to "design the environment for reliability" is what 13.7k engineers are endorsing here. #CodingAgent# #AIEngineering#

🔬 Can AI actually "discover" science? Tested on 90 real Nature papers, even the strongest agent beat SOTA on under 20% of tasks. Title: NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers? URL: 📋 Overview Using 90 tasks drawn from Nature-family journals across six domains, this work tests whether AI coding agents can reproduce and surpass published SOTA. An "information firewall" hides the original method, and NatureGym builds per-task containerized environments for strict, unified evaluation with web search disabled. 🎯 The problem it solves Prior agent-on-research benchmarks were fragmented and low in credibility. This work forces genuine discovery instead of reproduction and creates a common ground for cross-domain comparison. 📐 Methodology Tasks are scored with a SOTA-normalized relative gap (g > 0.1 to surpass, g >= 0 to match), comparing across 81 metrics. Ten agents are evaluated across three harnesses under a 4-hour budget with task-based GPU allocation. 📊 Results ・Even the top agent, Claude Opus 4.7, surpassed SOTA on only 17.8% of tasks and matched on 47.8% ・45.5% of successes were just translating a science task into supervised prediction; genuine domain reasoning was only 8.3% ・Failures were driven by wrong method choice (45.1%) and insufficient compute (24.4%), with task misunderstanding at only 3.1% ・Interdisciplinary tasks consistently scored worse A sober snapshot: today's agents are good at "translation" but weak at real "discovery." #AI4Science# #CodingAgents#