Give your agent an interpreter, not tools: 19 min to 2 min on the same model
Itz_hany · reddit · 2026-11-06
Checksum shares their "Code Mode" practice: instead of having agents call tools one by one, give them a stateful code interpreter so tools are invoked as Python functions. With the same model, test generation went from 19 minutes to 2, and cost per test from $1.67 to $0.25.
Two levels:
- Loop level: provide a REPL-style interpreter (e.g. Monty by Pydantic AI) so the agent can store results in variables, parse them, and chain calls in one shot.
- Task level: make the whole task a coding exercise — e.g. hand the agent the full HTML page instead of browser-tool snapshots (browser-use's browser harness), which they found faster and more accurate than Playwright MCP.
A side benefit: successful runs can be cached as reusable scripts by design.
More from coding & agent
- Team's 24/7 infra-monitoring agent spots anomalies 'before the fire,' user says — antoine_chaffin · 2026-11-06
- Delvetown: a human-agent "post-scarcity" web community, plus agents emailing researchers unbidden — Miles_Brundage · 2026-11-06
- Every team demos Codex and Claude workflows for running AI as your OS — every · 2026-11-06
- Hyperagent launches Sizzle, an agent that reads PRs and auto-produces polished demo videos — aakashgupta · 2026-11-06
- Vercel publishes CLI design guidelines for humans and AI agents — JohnPhamous · 2026-11-06
- Dev builds a sculpting tool for Apple Vision Pro using Codex — OpenAIDevs · 2026-11-06