Improving Deep Agents with harness engineering
Self-verification, tracing, and context optimization moved LangChain's coding agent from Top 30 to Top 5 on Terminal Bench.
AI Daily Recap
36 articles · 4 categories
The finishable daily brief
What happened in AI — Oct 13, 2026
Tuesday, Oct 13, 2026
36 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- Gemini 4 Argon launches on DeepMind and Google Cloud, with a safety write-up.
- LangChain: harness engineering took its coding agent from Top 30 to Top 5 on Terminal Bench.
- LangChain evals: GLM-5 and MiniMax M2.7 match closed frontier models on core agent tasks.
- Anthropic replaced its source control with an in-house system after a tenfold rise in merges per day.
- Modal keeps RL stable across mixed-precision, mixed-accelerator samplers.
- GitHub's secret classifier cuts median scan time by 86%.
Google launched Gemini 4 Argon on DeepMind and Google Cloud, but the more actionable news for builders came from agent teams quantifying what works: LangChain's harness changes lifted a coding agent from Top 30 to Top 5 on Terminal Bench without a model change.
Open models are closing the gap on agent tasks, and the infrastructure underneath is being rebuilt for agent volume, from Anthropic's new version control system to Modal's mixed-hardware RL.
Agent harnesses, evals, and production builds 5 items
Teams are showing measured gains from harness changes and eval-driven builds rather than model swaps.
Open Models have crossed a threshold
LangChain's evals say GLM-5 and MiniMax M2.7 now match closed frontier models on file operations, tool use, and instruction following, at lower cost and latency.
Coinbase Built, Evaluated, and Post-Trained a Fraud Agent
Coinbase details how it built, evaluated, and post-trained a fraud-detection agent.
Search agent traces by behavior with AI Search in LangSmith
LangSmith AI Search finds traces by describing the behavior in natural language instead of filtering on fields.
Quoting Geoffrey Huntley
Geoffrey Huntley's fix for agent drift: enforce diagnostics-library use with a pre-commit hook on the git diff, not AGENTS.md edits.
Gemini 4 Argon launch 3 items
Google shipped Gemini 4 Argon across DeepMind, Google Cloud, and robotics, with a safety write-up the same day.
Gemini 4 Argon: Built to solve your hardest problems
Google DeepMind introduces Gemini 4 Argon, pitched at its hardest-problem workloads.
Gemini 4 Argon available on Google Cloud, built to help solve your business’ hardest problems
Gemini 4 Argon is available on Google Cloud for enterprise use, tied to the Gemini agent in Gemini Enterprise.
Developing Gemini 4 Argon responsibly and safely
DeepMind publishes the safety and responsibility approach behind the Gemini 4 Argon release.
Training and developer infrastructure 3 items
Infrastructure posts focus on stable RL across mixed hardware and on scaling source control for agent-driven commit volume.
Toward RL on mixed-precision, mixed-accelerator samplers
Modal trained an agentic-coding model with fully asynchronous RL on mixed-precision, mixed-accelerator samplers; off-policy correction kept training stable and on par with homogeneous samplers.
Redesigning source control at Anthropic
Anthropic migrated to an in-house version control system in five months after merges per day grew more than tenfold.
GitHub’s model for secret detection: Finding more secrets with less noise
GitHub's context-aware secret classifier raises weighted acceptance by 11.5% and cuts median scan time by 86%.
Open-weight and frontier model moves 2 items
Open-weight agent tooling from Alibaba and a new Meta frontier model add competitive pressure on closed labs.
Alibaba's Qwen team open sources QwenGyre to train AI agents that run for hours
Alibaba's Qwen team open-sources QwenGyre to train agents that run for hours.
[AINews] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model
Latent Space covers Meta's Muse Wonder models and what they mean for MSL as a frontier lab.
You are caught up for this edition