Went on one of the biggest AI channels in the world to reveal two things: 1) How we're killing consulting. It needed to be taken out back, and rebuilt from the ashes. 2) The MetaHarness - an AI-native workflow + SDLC we used to 10x productivity Def watch the full video (@DavidOndrej1/@dan_zakon crushed it), but here are the highlights: 1) How to make a slow company AI-native - Meet the company where it actually is: systems, people, and where they sit on the AI journey. - A 2026 consulting business for Fortune 2000s has to be a full-stack partner, not a deck shop. - Old model: consultants take a problem, sit on it for months, deliver a 150-page deck, say peace, leave the client to implement. - Post-AI: you have to touch the whole chain. Strategy, change management around the tech, and forward-deploying into the company. - For a lot of Fortune 500s, the first move is teaching the C-suite Claude CoWork/Claude Code or Codex. - The job is to drive friction as close to zero as possible. 2) How we're killing & reviving consulting - Me and @ArmanHezarkhani did not dream of being consultants. - The vision: modern-day Bell Labs - My bar: if I'm not home with my wife and daughter, the work has to be worth it. Worth it = frontier problems with the smartest, most driven people you have worked with. - Arman read The Innovators in school (Babbage and Ada Lovelace through mobile revolution). The book’s lesson on innovation centers: Bell Labs / the transistor, DARPA / the internet, Xerox PARC. The raw materials are always brilliant people given autonomy, plus direction, distribution, and resources. - We did not want to raise hundreds of millions on day one. Helping the biggest companies solve their hardest AI problems cashflows the business, tells us what actually hurts, and supplies the materials for an applied AI lab that incubates consequential post-AI tech. - Consulting is a dirty word. It's growth & innovation as a service. - Classic consulting pitch: go in, problem, the answer is often firing people, deck, walk away, never see the fruit. That is why great technical talent never went there. - Reposition: messy, important problems, no answer, an A-team for innovation and implementation, actual ROI. - Consulting by-design gives engineers diversity of work without job-hopping. “ADHD is satisfied, same company.” 3) Singleplayer vs. multiplayer AI - Single-player: swap Google for GPT/Claude on the same query. Zero behavioral change. Still valuable. - Enterprise version of that: sign a $5–10M token agreement with OpenAI or Anthropic, get secure Codex/GPT or Claude Code/Co-work, put SOTA models in front of people, teach daily use. Most companies should start here. No compounding. - Multiplayer: reinvent a horizontal process so leverage hits a whole function or the whole company. - Examples: an SDR/sales agent that gives sellers back time from logistics so they talk to clients. Data engineering when an “AI problem” is actually a data-readiness problem. - Exponential, messy, painful value is multiplayer. 4) Breaking down the MetaHarness - Engineers have always been methodical about files. Knowledge workers are playing catch-up. Tenex is taking it further: extremely prescriptive context for coding agents. - Old SDLC: human-to-human coordination via Agile ceremonies (standups, retros, a calendar full of meetings). - New SDLC: humans, other humans, and agents. Agents are the main character doing the work on the ground. - Written-first culture matters more than ever. Markdown briefs humans and agents. - Agents can hold more coherent information than humans, so you can write bigger blueprints further in advance than Agile ceremonies ever allowed. - Concrete markdown artifact types on every project. Agents pull the right context at the right moment and keep it in sync. Dan calls it a “software machine.” - The problem the harness is solving: agents give speed that did not exist. Longer tasks increase entropy and diversion from plan. Minimize entropy without giving up speed.

Couldn't help myself. Forward deployed engineer or AI engineer sounds dry as cardboard, so I asked my team to come up with the most creative job titles describing what they do... The current list: - Transformer tamer - LLM whisperer - Hallucination exorcist - Token daddy - Weights guy - Human out of the loop - Latent freak - Gradient guru - RL gym junkie - Inference dealer - Nothing-but-the-harness - Benevolent prompt injector - The metaharness - Context window's context window Who can top this?

New video: @leonvz demos Omnigent running Claude Code and Codex in one open source meta-harness. He plans a 3D game, forks into Debby for a multi-model debate, then hands off to Polly to delegate the build. Policies, MCP, and prompt injection guardrails too. 👇 🎥 Watch: #Omnigent# #AIAgents# #OpenSource#

Check out this hands-on walkthrough of Omnigent from Daniel Beach: what a meta-harness actually is, without assuming you already live in this stack. The blog covers installing Omnigent, connecting Claude Code and Copilot, and the pieces that matter in daily work. 🔹 One common layer over your coding harnesses, plus a Web UI next to the terminal 🔹 Policies for cost and safety, including asking before git and bash 🔹 Polly for multi-agent planning and research 👉 Read the full post: #Omnigent# #AIAgents#

Nimble now runs natively inside Omnigent, Databricks' open-source agent meta-harness. See en example of what that looks like with our new cookbook: an agent that decides which API to use, then ships the integration. A Nimble Web Search Agent researches each candidate against one shared schema, returning per-field citations and confidence. A Claude head and a GPT head debate the evidence. A coding agent implements the winner on a reviewed branch. Cookbook: Docs:

📣 Hey Paris: Join us on Sept 23 for the next Open Lakehouse + AI meetup! Aravind Segu and Edwin He will cover why to use a meta-harness: collaborate on your work, exercise control, and choose your coding agent harnesses. The talk includes a demo of Omnigent’s workflow, including @opentelemetry traces in tools like @MLflow. 🗓️ Wed, Sept 23 | 6:00–9:30 PM GMT+2 📍 La Fondation, Paris 🎟️ Register: #Paris# #Omnigent# #OpenSource#

From our recent Omnigent & Multi-Agent Coding Meetup in NYC, @dennylee (@databricks) gives a primer on why a meta-harness matters. The talk covers: 🔹 Why coding agents and harnesses do not interoperate 🔹 How a meta-harness sits above Claude, Codex, Cursor, and similar tools 🔹 Multi-agent debates, where agents critique and improve each other’s code 🔹 Sharing a session so a teammate can view, comment, and run it if you grant them rights 🔹 Policies that require approval for file and shell operations 📽️ Watch now: #Omnigent# #AIAgents# #NYC#

Must-read paper from Google on self-improving agent harnesses. If you auto-optimize your agent's harness, your eval score can go up while the agent gets worse on real tasks. This paper shows how to prevent that. Of five harness-evolution methods compared on agentic workspace tasks, RRSI scored the lowest on the tasks it evolved against and highest on all three out-of-distribution benchmarks. Automated harness evolution proposes edits to prompts, control flow, tools and memory, keeps the ones that raise the score, and repeats. The authors show this overfits the training tasks. Meta-Harness reached 93.0 on the Harvey LAB evolve split but gained only 0.3 to 1.5 points on JobBench, GDPval and APEX-Agents. RRSI adds regularization on both sides of the loop. The proposer gets an edit budget that shrinks over time and is pushed toward directions it has not tried. A critic rejects benchmark-specific edits, and a pruner removes edits that are too small, too costly or no longer useful. RRSI scored 90.5 on the evolve split and gained 3.5 to 4.7 points on the three held-out benchmarks. In the ablation, unregularized evolution used 3.80M tokens per trial against 2.42M for RRSI. With Gemini 3.5 Flash, RRSI raised Terminal-Bench 2.1 from 64.6 to 78.7 and carried a 2.2-point gain over to SWE-bench Verified. Paper: Chat with Paper:

Things to try with Jev right now: 1. LLM-as-a-Judge evaluation 2. Routing for agent harnesses harness 3. Scaling agent orchestration by enabling smarter subagent creation with SOTA classification capabilities 4. Enhance dynamic harness generation where structured outputs are key The first three deliver insane ROI in cost and efficiency. For harness engineering, I'm using it as a smarter router for my meta harness. More on this soon. But honestly, I see other cool applications for improving tool calling and other context-engineering aspects of agents with this model. The fourth is something new I am currently testing but has huge potential to disrupt and enhance agent orchestration in new ways. All in all, this feels like an important primitive for improving your agents. If there is interest, I will write more about this in the coming days and share full guides. Let me know.

Harness optimization is getting real receipts. AutoSaddler is offline harness learning from agent failure traces. Not another prompt tweak loop. It patches prompts, tools, and middleware as code. Then it keeps updates that survive a held out set. On the test sets (Pass@1): GAIA2: 53.0 → 62.0 (+9.0) SWE-Bench Pro: 37.3 → 46.9 (+9.6) Terminal-Bench 2.0: 40.0 → 50.0 (+10.0) That TB2 number also clears the expert tuned Terminus KIRA at 47.5. Kill generalization aware selection and GAIA2 falls to 50.6, under the default agent. Deep diagnosis and structured patches help. Dev set filtering is what stops the harness from overfitting the mini batch. On GAIA2, Figure 1b, about 147 leveraged traces to the best dev score vs about 1,400 for Meta-Harness.