The Hidden Primitive Behind Coding Agents: Verifiers Must Sit Outside the Writer's Context
bibryam · x · 2026-09-08
AgentField co-founder & CTO Santosh Kumar Radha published part 2 of his harness orchestration series, arguing the real boundary of coding agents (Claude Code, Codex, Gemini CLI) isn't call-site parameters but the runtime 'membrane': files read, tools used, verifiers consulted, and side effects left behind.
Key points:
- The workspace is the largest prompt: before the first output token, agents scan the working directory, README, recent commits, test files, and dependency manifests. System + user prompts are only 2,000–4,000 tokens, while workspace reads contribute far more context.
- A verifier the agent can see shapes its behavior and ceases to be an independent contract. Acceptance checks should be placed outside the writer's context, tools, and workspace.
- The piece dissects the boundary surface along four engineer-shapable dimensions: the startup workspace, boundary drift at runtime, visible vs. invisible verifiers, and the blast radius a team can afford to undo.
More from coding & agent
- ReCache cuts agentic KV memory 92.43% and speeds first token 3.655x with minimal accuracy loss — techNmak · 2026-09-08
- ReCache reuses tool-schema KV states, cutting agent memory 92% with little accuracy loss — techNmak · 2026-09-08
- Weaviate's year of agent memory: why store-everything RAG breaks three ways — victorialslocum · 2026-09-08
- A simple Next.js app built 2 months ago already has 48 vulnerabilities — haydendevs · 2026-09-08
- GPT-6 Astra 'feels like AGI': $1,000 hands-on across real client work, gaming and computer control — PrajwalTomar_ · 2026-09-08
- Google open-sources Artemis, an Android-automation agent scoring 99%+ on AndroidWorld — Shruti_0810 · 2026-09-08