Operate an AI agent beyond the demo
Follow a proposed action through authorization, interrupted execution and an independent evaluation.
For practitioners who already work with services and APIs and want to make an agent’s actions understandable after a failure.
Before you begin
Experience with software services, APIs and operational changes. The first read refreshes the tool-call cycle; it is not a programming introduction.
What you will be able to explain
Draft a bounded operation and evaluation plan that separates permission, execution state and independently checked service results.
Read it, then decide what to try.
Follow the hypothetical cases as a design review. The two supplied Python examples use synthetic inputs; reading them does not require access to a live service.
Optional exercises: setup and scope
Python 3 is enough for the small tool-call fixture and standard-library grader. A full evaluation is separate work: it needs an isolated service or fixtures, synthetic data and credentials that cannot reach production.
- Software-service, API and operational-change experience is assumed.
- Use the supplied examples to inspect the mechanism; they do not install an agent or implement your service’s controls.
Start at the beginning, or where you left off.
This is a focused operating-design tour, not a deployment course. Follow Next in this tour; other lesson-series links are separate routes.
No progress is stored. Every reading stays open; bookmark an article’s tour link if you want to return to it.
Agents and tool calling: follow one complete exchange
Separate the model’s request, application validation and returned tool result.
Look for this: Keep the proposed call, application decision and returned result distinct. A synthetic lookup is not a real service observation.
Read step 1 →How production agents handle stale approvals and uncertain writes
Bind approval to the current target and distinguish an applied action from a useful service outcome.
Look for this: Distinguish an approved change, an applied change and a useful service outcome. An unknown write is neither a failure nor a success.
Read step 2 →Recovering an interrupted agent skill
Recover an interrupted skill from durable operation evidence instead of replaying its happy path.
Look for this: Identify what must survive a restart and which destination check happens before another write.
Read step 3 →Testing an incident-response agent against the service it must recover
Keep evaluator evidence separate and test failures, unknown results and useful completion before expanding a role.
Look for this: Ask whether an independent reviewer can explain the verdict without trusting the agent’s completion message.
Read step 4 →
Could another engineer continue the work?
Use these questions to find your next useful read, not to award yourself a score.
- Name the approved target and state, and explain what should happen if either changes before execution. Revisit the explanation →
- Show how an interrupted operation remains unresolved until destination evidence supports the next step. Revisit the explanation →
- Separate agent claims from evaluator observations, retain failed and unknown trials, and name who can stop the trial. Revisit the explanation →
Use these checks to draft one operation’s permitted scope, recovery states and independent success criteria. Where the destination offers no safe reconciliation, record that limitation rather than promising duplicate protection.
Finishing the reading is not authorization to operate production or evidence that an agent is ready for it.