What Is an Agent Harness? The 8 Shared Components Behind Claude Code, Codex, and Antigravity
Saboo_Shubham_ · x · 2026-10-30
The author frames 'agent = model + harness': the model reads context and picks actions, while the harness is all the code turning that choice into work. Claude Code, Codex, Antigravity, OpenClaw, and Hermes Agent are all harnesses.
Every harness shares 8 components:
- The loop: send context to the model, run requested tools, feed results back, decide when the task is done
- Tools & sandbox: terminal, files, browser, apps — isolated per task so bad commands stay contained
- Subagents: parallel side tasks with fresh context windows, returning only short summaries
- Context management: decide what the model sees each turn, load skills/files on demand, summarize old history before the window fills
- Memory: progress notes, git history, and skills on disk so new sessions resume where the last stopped
- Guardrails: hold risky actions (emails, payments, deletes) for approval; keep real credentials outside the sandbox so prompt injection has nothing to steal
- Verification: run tests, browser checks, or a separate grader before the agent can call it done
- Observability: trace every model/tool call, catch repeated failed edits, push the agent to try new approaches
More from coding & agent
- Your GitHub contribution graph is the wrong metric to move in the AI era — bilawalsidhu · 2026-10-30
- Annotake: open-source Mac tool sends only changed frames to multimodal models — Express-Phase1532 · 2026-10-30
- Teams of agents vs one executor with skills: Reddit debates the real technical difference — ladybawss · 2026-10-30
- Astra inside Codex harness shows 'AGI-ness' other frontier models lack — tszzl · 2026-10-30
- AI agent performs brain surgery on a live mouse: Claude picks drug, injects 250nL to treat epilepsy — BorisMPower · 2026-10-30
- CentOS creator builds Ginux kernel in 11 days: Codex wrote 139,000 lines of Rust — JFPuget · 2026-10-30