Direct the AgentArrows or space to advance · F for fullscreen · ⌘P to save as PDFSessions

The First Build · Session 03 · Dates TBD

Direct the Agent

Plan the feature, test the system, verify the outcome

Ship AIPhoenix, ArizonaPress → to begin

Ship AIDirect the Agent01/20

The agent is only as reliable as the evidence it can see.

A confident answer is not evidence. A diff, a passing check, a browser journey, a database row, and a clean log are.

Ship AIDirect the Agent02/20

Tonight

  1. 01The plan–execute–observe–iterate loop
  2. 02What the agent owns and what the human keeps
  3. 03Give the repository durable memory
  4. 04Scope one feature with acceptance criteria and edge cases
  5. 05Use checks and tests as executable specifications
  6. 06Verify the browser, API, data, and logs together
  7. 07Review the diff and leave a clean checkpoint
Ship AIDirect the Agent03/20

Where we are

  1. Session 01From Idea to Build Plan
  2. Session 02Make It Real
  3. Session 03Direct the Agent
  4. Session 04The First Build Hackathon

The product is live and its core data path works. Tonight the agent gets the context and evidence it needs to change that system without guessing.

Ship AIDirect the Agent04/20

Act one

Direct the loop

Agentic engineering is not a longer prompt. It is a loop where the agent can act, see what happened, and revise against a finish line.

Ship AIDirect the Agent05/20

The agentic loop

until done
  1. 01Plan from the repository and the outcome
  2. 02Execute with files, shell, browser, and APIs
  3. 03Observe checks, behavior, data, and logs
  4. 04Iterate on the evidence — or stop

The magic is not any one step. It is the feedback between them and a definition of done that prevents endless motion.

Ship AIDirect the Agent06/20

Prompting asks for code. Directing asks for an outcome.

One-shot prompting

  • A snippet answers the message
  • You carry context between files
  • You run the commands
  • You paste the error back
  • Done means it looks plausible

Directed agent

  • A feature satisfies criteria
  • The repository supplies context
  • The agent runs approved tools
  • Failures steer the next action
  • Done means the evidence passed

The human still owns scope, tradeoffs, access, and the decision that the evidence is sufficient.

Ship AIDirect the Agent07/20

Keep the product decisions human

The agentYou
PlanPropose files, steps, and risksApprove scope and tradeoffs
BuildImplement the agreed sliceAnswer product ambiguity
CheckRun the test gauntletChoose sufficient evidence
AccessUse granted toolsControl secrets and authority
ShipPrepare a clean changeOwn the release decision

Autonomy is scoped authority, not permission to decide what the product should become.

Ship AIDirect the Agent08/20

Exercise: give the repository memory

~/your-product

Create project instructions from this repository

stack + architecture summarizedcommands copied from package.jsonsecurity rules and protected files nameddefinition of done written

Write durable project instructions from facts the agent can verify: architecture, conventions, commands, risk boundaries, and what must run before a change is complete.

Do not write flattering adjectives about the codebase. Write the facts a new agent would otherwise rediscover or get wrong.

Ship AIDirect the Agent09/20

Act two

Turn the feature into evidence

Acceptance criteria say what the user can observe. Tests and checks make enough of that promise executable for the agent to steer by.

Ship AIDirect the Agent10/20

From intent to a verified change

  1. OutcomeName the user and the job that becomes possible.
  2. CriteriaWrite observable success, failure, and permission cases.
  3. Fail firstRun a check that proves the feature is not there yet.
  4. ImplementChange the smallest file set that can satisfy it.
  5. PassRerun the focused check, then the wider gauntlet.
  6. JourneyComplete the feature as a user on the real interface.

A failing check before the change proves the check can see the behavior. A passing check afterward then means something.

Ship AIDirect the Agent11/20

Each check catches a different lie

QuestionEvidence
TypesDo the data shapes agree?typecheck
LintAre there known unsafe or broken patterns?lint
BuildCan production compile this?production build
UnitDoes the rule behave at its boundaries?focused test
BrowserCan a user complete the journey?end-to-end check
DataDid the correct outcome persist?row inspection

No single layer proves the whole feature. Choose the smallest stack that covers the actual risk.

Ship AIDirect the Agent12/20

Live build: let failure steer

~/your-product

typecheck → lint → test → build → browser

focused test failed before implementationagent implemented 5-file planall automated checks passedbrowser caught missing empty state — fixed

Start with one failing criterion. Let the agent implement and rerun checks. Then use the browser as a stranger and feed the first mismatch back as evidence, not as a guess about the cause.

A browser check catches the parts a green build cannot see: focus, copy, timing, empty states, and the actual sequence of actions.

Ship AIDirect the Agent13/20

Act three

Verify the outcome

The feature crosses browser, server, data, and deployment. Verification should cross the same surfaces.

Ship AIDirect the Agent14/20

The end-to-end evidence chain

  1. BrowserPerform the core journey and inspect the visible states.
  2. NetworkConfirm the intended request, response, and error behavior.
  3. DatabaseInspect the saved row, owner, defaults, and timestamps.
  4. LogsCheck for hidden exceptions, retries, and sensitive output.
  5. DiffReview scope, security, and files the plan never named.

Tests check claims about code. This chain checks the outcome the product created.

Ship AIDirect the Agent15/20

Interrupt the agent when the evidence changes the plan

  • Scope expandsThe agent wants to refactor adjacent systems to finish a narrow feature.
  • Authority expandsIt needs production data, a new secret, a migration, or an external write you did not grant.
  • Risk changesThe work now touches auth, payments, destructive operations, or irreversible data.
  • The check is weakThe test passes without exercising the behavior you actually care about.

Persistence is useful inside a stable contract. When the contract changes, stop and make the new decision explicitly.

Ship AIDirect the Agent16/20

The reviewed checkpoint

~/your-product

review diff → verify live path → commit

unplanned dependency removedpermission case and empty state verifieddatabase defaults correct; logs cleanwrote 03-agentic/verification-log.md

Read the diff for unintended scope and security changes, perform the live journey once more, then record the commands and observations that justify the checkpoint.

Save the brief, plan, tests, verification log, and reviewed diff in 03-agentic/. The hackathon starts from this operating loop.

Ship AIDirect the Agent17/20

What the loop gives you

BeforeNow
ContextRe-explained in every promptRepository instructions
ScopeA feature-sized wishCriteria and a reviewed plan
FailureA reason to try another promptEvidence that points to the next step
DoneThe code looks rightThe user outcome is verified

This is the reusable method: plan, act, observe, revise, and stop when the agreed evidence passes.

Ship AIDirect the Agent18/20

Failure is a clue when the loop can see it.

Ship AIDirect the Agent19/20

Next

The First Build Hackathon — Session 04, Workuity Biltmore

Bring the repository, one narrow problem, and the evidence loop. The capstone ends with a live product journey, a frozen deploy, and an honest account of what the agent got wrong.

shipai.club · Free, always

Capstone page
Ship AIDirect the Agent20/20
01/20