Verification patterns for tool-calling and browser agents: separate actor and verifier
InsideDebt6345 · reddit · 2026-09-19
The author shares verification patterns for agents that write code, call tools, or drive browsers, aimed at stopping hallucinated "done" claims:
- Separate actor and verifier: the verifier is deterministic code with no LLM, checking outputs against clear rules before any real action runs.
- Force structured outputs: anything triggering side effects must be JSON with checkable fields and IDs, not free-form text, making validation and retries easier.
- Evidence and caps: store one evidence artifact per side effect (log, ID, snapshot); cap tool calls, cost, and runtime in the runner.
The author says this prevents agents from confidently claiming success when nothing changed, and links a 30-minute implementation guide. They close by asking how others handle verification layers and what riskiest agents people are running.
More from coding & agent
- One prompt, under $1: coding agent builds a full Django to-do app in 5 minutes — Al_Grigor · 2026-09-19
- One Generic Video Tool or Per-Provider Tools? Debating the Agent Tool Layer — ExcitingBison4616 · 2026-09-19
- Readback: free MIT VS Code extension reads Claude Code replies aloud via Speechify — shauntrennery · 2026-09-19
- GitHub Next open-sources LocalJev, a local Jev-compatible API built on oMLX and DiffusionGemma — gaganghotra_ · 2026-09-19
- WebMCP benchmark: Jev + Mercury 2.5 solves 100% of tasks at 112x lower cost than GPT-6 Astra — hardimanjames · 2026-09-19
- Musecases Launches: A Community-Voted Prompt Library for AI Agents — ChrisUniverse · 2026-09-19