@t3chn
Nothing here yet.
Nothing here yet.
3d ago · 7 min read · Most agent tooling mistakes coordination for reliability. It gives you more roles, more agents, more orchestration, and more shell theater. The demo gets more impressive. The system does not necessarily get easier to trust. That tradeoff used to be t...
Join discussionMar 23 · 5 min read · Last week, our pipeline produced a proofpack with decision: HUMAN_REVIEW. The contract had 10 acceptance criteria. The engineer agent created all the new files, build passed, tests passed, three independent reviewers ran. Everything looked correct — ...
Join discussionMar 20 · 5 min read · I maintain 12 Claude Code plugins. Each has a project.intent.md -- a structured spec that tells the agent what the project does, what it doesn't do, and who it's for. The agent reads it at the start of every task. Last week I ran a reverse diff -- co...
Join discussionMar 18 · 5 min read · You can review an AI diff with three models and still have zero ground truth. They'll tell you what looks "reasonable", not what's correct. The failure mode isn't "bad code". It's unfalsifiable intent: the requirement never became something you can r...
Join discussionMar 17 · 6 min read · The morning digest surfaced Leanstral -- Mistral's open-source agent for formal verification in Lean 4. A mixture-of-experts model (119B total, 6.5B active per token) that scores within 80% of Claude Opus on the FLTEval theorem-proving benchmark at a...
Join discussion