In short
In the agentic PDLC the engineering lead writes the context file every coding agent reads and builds each bolt from a story file. The lead puts every number the product acts on in tested code, enforces every limit in a tool's signature, and wires the harness that blocks a merge when a slice falls below its bar. The lead ships a slice a day, letting the model type and never letting it hold the boundary.
The engineer across the four phases
- Not on the clock: reads the brief, starts nothing
- 1 · Preparethe context file
- 2 · Slicea story file per bolt
- 3 · Floorexact code first
- 4 · Layermodel calls, a checker
- 5 · Gatethe limit in the signature
- 6 · Harnessin CI, in cost order
- 7 · Shipa bolt a day, behind a flag
- 8 · Operatecache, route, trace, ledger
An engineer who opens a branch in P0 is the most expensive habit in agentic delivery. The work is P2: floor first, then the model, then the checker.
In this lesson you'll learn:
- the engineering lead's eight steps, and why P0 is deliberately not on your clock;
- the three things you own that no model may: the floor, the boundary and the merge check;
- how to use coding agents so they make you faster rather than busier.
Sound familiar?#
- The agent re-implemented a calculation inline, confidently, and it was wrong by eighteen dollars.
- "Review your answer before returning it" was added to the prompt; scores did not move and the bill rose.
- The evaluation job posts a comment on every pull request, and on the Thursday before the pilot someone clicked past it.
Each is a boundary held by a request instead of by code. The engineer's job is to move each one into code.
What changes for an engineer?#
You stop treating a prompt rule as a control, and start shipping a slice a day. Coding agents make the typing cheap; they do not make it correct. What you own is everything around the typing: the context the agent reads, the file it builds from, the deterministic floor under it, the limits it cannot cross, and the check that decides whether what it built may merge.
Your eight steps#
P0 · Frame: not on the clock#
You read the brief and start nothing. An engineer who starts building in P0 is building the prototype the requirements will later be written around.
P1 · Design & Spec#
1 · Prepare the context file every coding tool reads: stack, context layers by path, conventions, commands that have actually been run, a never-touch list. An hour's work, and the biggest quality lever you have. AIDD's five habits
P2 · Build & Prove: the phase you lead#
2 · Slice: build each bolt from a six-part story file, never a chat thread. 3 · Floor: every number that gets acted on is a function with a unit test. SkyWays' prompt returned $80 where the ledger said $62. 4 · Layer: the model calls on top, with an independent checker after the expensive ones. 5 · Gate: caps and confirmation tokens in tool signatures, each with two tests seen failing first. 6 · Harness: the golden set in CI as a required check, in cost order, per slice. 7 · Ship: one bolt a day in dependency order, the shadow path behind a flag, the rollback rehearsed. P2 Build & Prove
Skeleton first with no model in it. Exact code early: it stands alone.
The plug lands before the gated write that needs it. The proof goes last.
A bolt that cannot be built alone was cut wrong. Say so before you start it.
P3 · Run & Learn#
8 · Operate: caching that hits, routing by complexity with a loop cap, a trace that redacts, an injection suite run weekly, and the effort-and-token ledger the product manager's cost number is built from. Why the bill is 4×
What is yours, and what is not#
| Yours to own | Not yours |
|---|---|
| The context file, and the story file every bolt is built from | The bolt cut: the architect's; you say whether each bolt can be built alone |
| The deterministic floor: every acted-on number is a tested function | The acceptance bar: the product manager derives it; you make it run |
| The boundary: caps and confirmation tokens in signatures | The golden set's contents and the judge rubric: QA's |
| The harness in CI, and the per-slice rule that blocks a merge | The cut-over and widening: you build the flag; the PM throws it |
| Build order within the cut, integrated the same day, and the ledger |
How to use a model in this role#
Use a model for the typing and the sweep, never for the boundary. It will write a correct function faster than you can, and just as happily write a cap into a prompt and report the cap as done. Let it draft inside something you wrote: a context file, a story file, a signature you already fixed. Every line that moves money, changes a booking or writes a trace row is read by a person before it merges.
Where you'll use it#
- On day one of a repository: the context file, before the first agent session.
- The evening before each bolt: its story file, so the chat window can stay shut.
- On every change to a gated tool: two named readers, and the over-cap and no-confirmation tests.
Why it matters#
Coding agents amplify whatever habits surround them. With a floor, a boundary and a harness, they produce more working software; without them, they produce more plausible software, faster, and the difference only shows in production.
Try it#
An agent wrote issue_credit(customer_id, amount) and the prompt says credits over $50 need approval. What do you change before this merges?
Show the answer
The signature, and two tests. Give amount a bound so a value over $50 raises unless a valid confirmation token, created only by a person's approval, is passed; write test_over_cap_raises and test_no_confirmation_raises, and see both fail before the fix makes them pass. Keep the sentence in the prompt so the agent behaves well by default. Band the tool as money, so the change needs two named readers.
Key takeaways#
- The model types; you hold the floor, the boundary and the merge check.
- Context file first, story file per bolt, and every acted-on number in tested code.
- Ship a bolt a day behind a flag, with the harness deciding what merges.
FAQ#
How do software engineers work with AI coding agents?#
By giving the agent what it needs to be right and taking away what lets it be dangerous: a context file it reads every session, a story file per unit of work, tested functions for anything exact, limits in tool signatures, and a harness that blocks merges when quality drops, then reviewing by risk.
What does an engineering lead own in an AI project?#
The context and story files, the deterministic floor, the enforced boundary in tool signatures, the evaluation harness in CI, the build order within the architect's cut, and the ledger of effort and tokens that the cost report is built from.
Should AI agents write production code?#
Yes, inside structure: from a story file, against a context file, on top of tested exact code, with limits enforced in signatures and a harness that gates the merge. Code that moves money or changes records is read by people before it merges.
What is a deterministic floor?#
The set of functions, each with a unit test, that do every exact thing the feature relies on (arithmetic, lookups, published rules) so that no number the product acts on ever comes from a model.
Apply it in your role#
| If you are… | Do this | The AI-augmented shortcut |
|---|---|---|
| A forward-deployed engineer | In the customer's codebase, start with the context file and one walking skeleton. A merged change on day one is your credibility. | Have a coding agent read the repository and draft the context file before your first meeting. |
| A product manager or FDPM | Give engineers story files with the spec lines and the done-when. A coding agent cannot ask you what you meant. | Ask a model to rewrite a ticket as a story file and list every question an engineer would still have. |
| A GenAI or agentic AI engineer | Your loop: exact code first, then the model step with its checker. The harness decides the merge, and caps live in signatures. | Run the coding agent from the story file with a failing test first, and let the test define done. |
Across the enterprise. Standard scaffolding (context file, story template, harness, per-call log) across repositories makes any engineer productive on any team's agent in a day.
The ten-minute workflow. Test-first with a coding agent:
Here is the story file for today's bolt: <paste>. Before any implementation, write the failing tests:
unit tests for every exact step, one golden-slice test for the best-guess step against its bar, and one
refusal test for every cap. Then implement until they pass. Stop and ask if the story file does not say
what "right" means for a step.Sources and credits#
| Idea | Origin | Source |
|---|---|---|
| The engineering lead's eight steps, owns and not-yours | Original: this manual | Engineering lead, end to end · Role: Engineering lead |
| The shift: stop treating a prompt rule as a control | Original: this manual | For leadership |
| The walking skeleton | Borrowed | Cockburn, A. (2004). Crystal Clear. Addison-Wesley |
| The SkyWays examples | Illustrative: a fictional airline | Journey: Engineering lead |