https://oddly.fyi/2026-09-03T12:00:00ZChaitanya Meesalahttps://oddly.fyi/writing/trustworthy-autonomy/2026-09-03T12:00:00Z

Trust is not a personality trait of the model. It is a property of the system surrounding it.

https://oddly.fyi/notes/what-does-good-mean-for-an-agent/2026-09-03T12:00:00Z

AI exposes how much of quality was previously implicit infrastructure.

https://oddly.fyi/writing/agent-harness-is-the-product/2026-08-29T12:00:00Z

The model supplies general intelligence. The harness turns that intelligence into dependable work in a particular world.

https://oddly.fyi/notes/skills-arent-policies/2026-08-28T12:00:00Z

Procedure, authority, and enforcement are different concerns.

https://oddly.fyi/notes/memory-should-accumulate-judgment/2026-08-21T12:00:00Z

Transcripts preserve history. Useful memory changes the next decision.

https://oddly.fyi/writing/rollouts-are-decision-problems/2026-08-18T12:00:00Z

A rollout reviewer should not merely detect anomalies. It should decide what the available evidence justifies doing next.

https://oddly.fyi/notes/context-vs-harness-engineering/2026-08-14T12:00:00Z

Context makes the world legible; the harness makes action dependable.

https://oddly.fyi/notes/why-agent-verifiers-fail/2026-08-07T12:00:00Z

A second model does not automatically create independent assurance.