Dev says self-verification tools are his biggest agent upgrade in months
Developer Daniel Lockyer shared lessons from months of testing various agent skills: letting the agent verify its own work was the highest-payoff improvement, mattering more than stacking prebuilt skill packs.
Confirmed
- Give the agent a small CLI for making API calls so it can actually check interface behavior
- Use Playwright to open pages and take before/after screenshots to verify UI changes
- Connect infrastructure providers via MCP so that after completing a code fix the agent can automatically watch the deployment, check metrics and spans, and confirm the change achieved the goal—otherwise heavy manual intervention is needed
- Concrete case: he asked AI to change some UI; after modifying the code, the AI invoked the verification tool itself to generate a screenshot, found the change caused unplanned side effects, fixed them, and attached before/after comparison images in the PR for human confirmation
- His conclusion: he prefers "native AI plus self-built verification features" over most off-the-shelf skill packs
Why it matters
- This experience shows the reliability bottleneck for agents often isn't capability but the lack of self-checking means; verification tools let AI discover and correct its own errors, reducing manual review cost
- Practices like before/after screenshots and metric observation offer a replicable engineering pattern for teams adopting AI coding
2026-10-07 ~ 2026-10-07 · 5 related posts
Primary sources
- Giving agents self-verification tools is the biggest win, says dev after months of testing — DanielLockyer ·
- AI Catches Its Own UI Side-Effect via Screenshot Verification Before Opening PR — DanielLockyer ·
- Dev prefers vanilla AI + verification over skill packs: wire MCPs so agents fix, deploy, and verify metrics themselves — DanielLockyer ·
- [source] Giving agents self-verification tools is the biggest win, says dev after months of testing — DanielLockyer · 2026-10-07
- Agent skill that pays off most: self-verification with Playwright before/after screenshots — DanielLockyer · 2026-10-07
- Hooking MCPs to infra lets agents fix code, watch deploys, and verify metrics — DanielLockyer · 2026-10-07
- [source] Dev prefers vanilla AI + verification over skill packs: wire MCPs so agents fix, deploy, and verify metrics themselves — DanielLockyer · 2026-10-07
- [source] AI Catches Its Own UI Side-Effect via Screenshot Verification Before Opening PR — DanielLockyer · 2026-10-07