talvinder

We tried adding traces after an agent pipeline was built. The missing evidence was already gone: input version, checks run, results and escalation. Define that receipt before the agent runs. If a failed check leaves no inspectable record, block promotion.

A demo turned a messy PDF into a narrated course in 90 seconds. Six weeks later, roughly 1 in 20 outputs was subtly wrong, so a person still read every line. My release rule now: measure silent errors on real inputs before removing review. Demo speed is not operating readiness.

Early Ostronaut agents sometimes chose the wrong API, malformed parameters or retried forever. We added validation before execution; it rejected 30% of bad calls. For each tool call, check tool name, schema, permissions, retry budget and idempotency before touching customer data.

In QA, Ostronaut turned a case-specific 1,000–1,200 calorie upper bound into a general weight-loss answer. It matched the source and was unsafe. We stopped using LLM judges for safety-critical facts and added rule-based domain gates. Test qualifiers, not just numbers.

Our first multi-agent pipeline burned through its OpenAI budget in two weeks. The cause was not model price; several agents were rereading the same source material. We cached shared context and cut cost 40%. Before switching models, trace duplicated context by agent and step.

Adding AI creates new operating work. Someone must own evaluations, evidence quality, model and prompt changes, provider failures, escalations and the boundary between suggestion and action. Put those owners in the product operating model before launch.

Test the reason a result exists, not only the row it produced. For every AI action ask: did the source exist, support the claim, permit this use, match the approved version, and return a destination receipt? Fail the test when any answer is no.

Give every agent a source policy, not a search box. Define allowed sources, maximum age, authority rules, duplicate handling and failure behaviour for the job. A news monitor and a research archive should not accept the same evidence.

Before an AI sends or writes anything, the approval screen should answer: what changes, who sees it, can it be undone, and what remains after a partial failure? If the operator cannot answer those four questions, do not enable Approve.

When an operator sees ‘Customer promised Friday,’ they should be able to open the exact recording span, confirm the speaker, inspect the interpretation and approve the resulting task. If any step is missing, keep it as a candidate.