Tag
All content about Harness, organized for fast scanning.
Recent developments in autonomous coding agents highlight a shift towards goal-driven loops that enhance their efficiency and effectiveness. Innovations include open-source tools for vulnerability patching and improved architectural frameworks that significantly increase developer productivity. Additionally, there is a growing focus on creating long-running agents capable of maintaining context and memory across sessions, addressing current limitations in agent performance.
Addy Osmani says coding agents may be shifting from prompt-by-prompt use to goal-driven loops that plan, split work, verify results, and repeat. His thread maps the core building blocks—and the token, security, and “exit condition” pitfalls. [https://x.com/addyosmani/status/2064127981161959567](htt…
Anthropic has just rolled out Defending Code Reference Harness, an open-source reference for using Claude to find and fix vulnerabilities. It includes guided Claude Code skills plus a sandboxed C/C++ pipeline with ASAN and gVisor, though the repo is unmaintained.
Basis shared how it reorganized its monorepo to better support coding agents, built around “canonical” context and a six-layer AGENTS.md and skills architecture. The company claims the changes drove 5x higher token usage per developer and 2.5x faster weekly commits.
Anshuman Mishra lays out a bottom-up recipe for agent training using a tiny text-to-diagram task. The key: start with a strict environment and reward loop, use SFT to learn valid actions, then apply RL to optimize behavior—and watch for reward hacking.
A recent article by Addy Osmani explores “long-running agents” that can keep working across sessions without losing state. It outlines the key architectural patterns and why today’s agents still stall when context, memory, and verification break down.
Joongwon Kim and coauthors argue test-time scaling for long-horizon coding agents depends less on more sampling and more on carrying forward useful rollout information. Their summary-based RTV and PDR methods boost results on SWE-Bench Verified and Terminal-Bench v2.0.