The hidden primitive behind Claude Code, Codex and Gemini: verifiers agents can't see
bibryam · x · 2026-09-29
AgentField's harness orchestration series (part 2) introduces the "membrane"—the boundary surface of a single harness that orchestrators actually control, as opposed to call-site parameters like provider, model, or maxturns. Key points:
- The workspace is the largest prompt: before writing, agents scan the directory, README, commits, and test files—system and user prompts are only 2k-4k tokens, while workspace orientation often accumulates 30k-60k tokens the orchestrator never explicitly placed.
- The article dissects the boundary along four dimensions: the startup workspace, the boundary that drifts during a run, the verifiers the agent can and cannot see, and the affordable blast radius.
- Core claim: a verifier the agent can see shapes its behavior and is no longer an independent contract; acceptance checks must sit outside the writer's context, tools, and workspace.
- The piece frames this as the shared hidden primitive behind Claude Code, Codex, and Gemini.
More from coding & agent
- StepFun co-founder proposes KITE: PD-separation-inspired training for scaling agentic LLMs — teortaxesTex · 2026-09-29
- LangChain team talk by Sydney Runkle and Victor Moreira is worth watching even if you don't use LangChain — js_craft_hq · 2026-09-29
- CS student asks how shipped agents handle confident-but-wrong actions on real systems — Professional-Mine681 · 2026-09-29
- WebBrain: open-source local browser agent with a 450M browser-specialized VLM — ButtercupLyn100 · 2026-09-29
- NVIDIA OpenShell tested: 10/10 secret leaks without it, 0/10 with default policy — but auto-approve leaked in 12/12 — No-Peanut-6988 · 2026-09-29
- Sonnet 5.5 one-shots a full $100K/month app in a single prompt — PrajwalTomar_ · 2026-09-29