A 4B Model Verifies a Coding Agent 70x Its Size at Just 4% of Steps, 9-13% Extra Cost
petrusenko_max · x · 2026-10-30
The author breaks down SWE-Verifier-style selective verification: a small 4B model checks only critical decision points (<4% of steps) for 9-13% extra API cost, acting as a verifier for an agent 70x its size. He shares his own WhatsApp DM agent implementation, with a diagram covering the full design: a pre-send review gate where the second model reads the last 10 chat messages, runs via subscription-based claude -p (no API keys), falls back to DeepSeek through Hermes with no tools, and runs alongside deterministic rules. Key principle: place the check at the irreversible action—the send—rather than auditing every step.
More from coding & agent
- AI revives formal verification: seL4 took 20 person-years, startups now claim it's practical — lemire · 2026-10-30
- YC experiment: 4 AI agents spend 11 minutes appointing a "sol" in Minecraft bureaucracy — aiamblichus · 2026-10-30
- Microsoft releases Mage-VL-Cowork, a vision-language model for computer-use and coding agents — microsoft · 2026-10-30
- WhatsApp outreach agent: one promo per person per 14 days, counted across all DM lanes — petrusenko_max · 2026-10-30
- Nova v1 'Flash' launches as an AI engineering teammate aiming to replace the IDE — Gauri_the_great · 2026-10-30
- Malicious Claude Code extension with SSH backdoor hit Cursor's marketplace, up to 5,000 users — cyb3rops · 2026-10-30