Agentic programming

The course, in order

Open a unit in the map to inspect its activities. Reading, guided completion, independent work and review remain distinct evidence.

1. Give an AI a verifiable programming task

   Separate the requested behavior, scope, and acceptance evidence in a coding brief.

   Open this unit
2. Supply relevant context without secrets

   Choose task-relevant repository context while keeping sensitive files out of requests.

   Open this unit
3. Write acceptance tests before trusting a patch

   Turn boundary requirements into executable checks and keep missing evidence visible.

   Open this unit
4. Inspect proposed changes against scope

   Detect unapproved file changes and separate edits from deletion intent.

   Open this unit
5. Use a failing test to guide an AI repair

   Demonstrate a defect before the patch and verify the same contract afterward.

   Open this unit
6. Validate an AI’s tool requests

   Restrict tool names and validate arguments before executing a model proposal.

   Open this unit
7. Keep consequential actions behind explicit approval

   Match an approval to its exact action and distinguish review from execution permission.

   Open this unit
8. Control an assistant’s repair loop

   Stop on success or exhaustion and surface failure rather than retrying indefinitely.

   Open this unit
9. Measure cost and inspect the action trace

   Account for token cost and preserve failures in a readable execution history.

   Open this unit
10. Accept results using evidence, not confidence

    Require current checks and review findings before accepting an AI-generated change.

    Open this unit
11. Turn a large request into a dependency graph

    Decompose a programming objective into independently verifiable tasks whose dependencies and acceptance evidence are explicit.

    Open this unit
12. Route work to the least costly capable model

    Choose a model from explicit capability, context, privacy, latency, and cost constraints instead of using one model for every step.

    Open this unit
13. Checkpoint agent state for safe resume

    Persist the minimum trusted state needed to resume a workflow without replaying completed side effects.

    Open this unit
14. Delegate work with explicit handoff contracts

    Assign bounded work to specialized agents and merge their findings without losing provenance or ownership.

    Open this unit
15. Run agent work concurrently without corrupting state

    Control concurrency through resource ownership, deterministic joins, cancellation, and bounded fan-out.

    Open this unit
16. Keep untrusted content out of the control plane

    Separate trusted instructions from repository, web, email, and retrieved content that may attempt to redirect an agent.

    Open this unit
17. Evaluate agent behavior by task and failure slice

    Build reproducible agent evaluations that measure task success, tool behavior, safety boundaries, and regressions across meaningful slices.

    Open this unit
18. Design retries that do not duplicate consequences

    Classify failures, apply bounded backoff, and make retried tool actions idempotent or explicitly reconciled.

    Open this unit
19. Operate agent workflows from structured traces

    Instrument agent runs with correlated spans, decisions, costs, artifacts, and redacted failures that support diagnosis without exposing secrets.

    Open this unit
20. Ship an accountable agent-assisted change

    Evaluate supplied version-linked release records and choose the first invalid required workflow stage. Then use the connected repository workspace to collect actual file, test, Git and reload evidence; the selectors alone do not execute a production workflow.

    Open this unit