New paper defines four conditions that turn an LLM into a coding agent
alex_verem · x · 2026-07-22
A new paper argues that a language model becomes a coding agent only when a harness provides four things: a runtime loop that interleaves reasoning, action, and observation; a tool interface to act on the environment; active context management; and execution-state tracking.
The paper traces the term harness back to the horse-tack metaphor and claims the agent is not the model itself but the surrounding infrastructure that connects, controls, and measures it. The screenshot shows the paper’s abstract framing this as a formal definition of agent harnesses for software engineering.
More from coding & agent
- Google says Gemini 3.5 Pro is in partner testing and Gemini 4 pre-training has started — eyishazyer · 2026-07-22
- Google’s Gemini 3.5 Flash Cyber is a restricted vulnerability-finding model for CodeMender — eyishazyer · 2026-07-22
- Google launches Gemini 3.6 Flash with lower token use, higher coding scores, and new pricing — eyishazyer · 2026-07-22
- Google’s Gemini 3.5 Flash-Lite pushes 350 tokens per second at $0.3 in and $2.5 out — eyishazyer · 2026-07-22
- Ask Sol to explain ExploitGym, and it starts cloning the repo and setting up firewalls — JensHonack · 2026-07-22
- Hermes agent user asks which cheap open-source model can handle long-running automation — OkCare6395 · 2026-07-22