SkyWaysThe agentic manual

The hard gate: the hand-off you can't skip in AI delivery

Most decisions in an agentic build should not stop anything. Three must, and all three sit at one hand-off. Knowing which is how you keep speed without buying an incident.

7 minutesIntermediateLesson 5 of 9Updated By

On this page
  1. Sound familiar?
  2. What is the hard gate?
  3. How to decide what is hard, step by step
  4. Where you'll use it
  5. Why it matters
  6. Try it
  7. Key takeaways
  8. FAQ
  9. Apply it in your role
  10. Sources and credits
All lessonsStart here

Getting started

  1. What is the agentic PDLC?
  2. How this tutorial works

Agentic PDLC fundamentals

  1. The evolution of the PDLC
  2. Why agentic AI projects fail
  3. P0 · Frame
  4. P1 · Design & Spec
  5. The one hard gate
  6. P2 · Build & Prove
  7. P3 · Run & Learn
  8. The eight loops
  9. The evidence pack

Methods decoded

  1. AI-DLC vs AIDD vs agentic SDLC
  2. What is AI-DLC?
  3. What is AIDD?
  4. What is the BMAD Method?
  5. What is spec-driven development?
  6. One lifecycle for every method
  7. How much process a change needs

Running delivery

  1. How to run an agentic AI project
  2. Bolts vs sprints
  3. A board for agentic work
  4. Cut delivery from months to weeks
  5. Review AI code by risk
  6. How accurate must an agent be?
  7. Prove the agent meets its bar
  8. Shadow mode and cut-over
  9. Why the AI bill is 4× the estimate
  10. Guardrails that hold
  11. The five governance gates
  12. Catch AI drift
  13. AI incident postmortems

By role

  1. For product managers
  2. For program managers
  3. For solution architects
  4. For engineers
  5. For forward-deployed engineers
  6. For QA
  7. For DevOps and platform
  8. For business sponsors
  9. For executives

Teams and organisation

  1. Team structure for agentic AI
  2. Measuring AI productivity
  3. The maturity model
  4. Rolling it out in 90 days

Practice

  1. The SkyWays case study
  2. The operating rhythm
  3. The workbench
  4. Twelve exercises

Interviews and careers

  1. What is an FDE?
  2. Six answer frameworks
  3. AI PM interview questions
  4. FDE interview questions
  5. GenAI engineer questions
  6. Agentic engineer questions
  7. AWS GenAI questions

In short

Of the four hand-offs in the agentic PDLC only P1 → P2, the hard gate, halts the build. The spec, the acceptance bar per slice and the authority budget are what everything downstream is built and measured against. Every other open decision runs alongside the build behind a placeholder, with a named owner and a date.

Hard or soft?the test for any open decisionAn open decisionCheap to reverse once building starts?noyesCan the build proceed behind a placeholder?noyesA named owner and a date?noyesDoes downstream work survive a change?noyesHARDsettle it before the phase closesSOFTruns beside the build, behind a placeholderOnly one crossing is hard1.P1 to P2: the spec, the bar and the guardrails2.Soft gates cross with a placeholder, an owner, adate3.Treating all four as hard is its own failure

Hard or soft? the test for any open decision

An open decision

  1. Cheap to reverse once building starts?
  2. yes
  3. Can the build proceed behind a placeholder?
  4. yes
  5. A named owner and a date?
  6. yes
  7. Does downstream work survive a change?
  8. yes
  9. SOFTruns beside the build, behind a placeholder
  10. any noHARDsettle it before the phase closes
Only one crossing is hard
  1. P1 to P2: the spec, the bar and the guardrails
  2. Soft gates cross with a placeholder, an owner, a date
  3. Treating all four as hard is its own failure

In this lesson you'll learn:

  • why the P1 → P2 hand-off is the only one that halts the build;
  • the four questions that sort any open decision into hard or soft;
  • how to cross a gate on credit without losing track of what you owe.

Sound familiar?#

  • The build is waiting behind eleven open questions, most of which could have been answered next month.
  • A decision that "could be changed later" turned out to be built into everything by the time anyone tried.
  • The gate was passed in a meeting that ended in "fine", and nobody can say what evidence was in the room.

Both failures, halting on everything and halting on nothing, come from treating every decision the same way.

What is the hard gate?#

The agentic PDLC has four hand-offs between its phases. Three are soft: work may cross with a placeholder, an owner and a date. One is hard: nothing enters P2 until three artefacts are signed. The home page and the simulator call it by its plain name, the sign-off.

ArtefactWhy nothing downstream survives without itWhat happens if it is missing
The eight-field specIt is what the coding agent reads, and a machine cannot ask what you meantThe agent invents an interpretation, consistently and plausibly
The acceptance bar per slice"It works" is no longer yes or no, so someone must say what share is enoughThe team ships at whatever number it got and calls that the bar
The authority budget and gate mapIt lists what the machine may do, with caps typed into signaturesThe tool accepts any amount, and the cap lives in the prompt

It sits there because P1 → P2 is the last point at which changing your mind costs a document rather than a rewrite. It is a one-way door; the others are two-way.

two-wayone-way
Mental model Reversibility is the hinge

How to decide what is hard, step by step#

Step 1 · Ask the four questions of every open decision#

List every decision that is still open, then ask four questions of each, in the order in the diagram. One "no" makes it hard. Can it be reversed cheaply once the build has started? Can the build proceed behind a placeholder? Is there a named owner and a date? Does everything downstream survive if the answer changes?

Step 2 · Settle the hard ones before the phase closes#

In a typical build three decisions come back hard, and each for a different reason:

Hard decisionWhy it halts
The AI-fit verdictIt fixes the shape of the product; changing it later means starting again
The autonomy level on money actionsIt moves money, and a regulator has a rule about it
The spec, the bar and the guardrailsEverything downstream is built and measured against them

Step 3 · Put a placeholder behind every soft one#

The rest (SkyWays had eight) run beside the build. What makes that safe is a real placeholder that lets the work proceed while the decision is measured: an interface layer in front of the framework, a stub that returns a file instead of a retrieval system, every call on a mid-tier model behind a gateway until the shadow run shows which slices need more. Each has an owner and a date on which it will be settled.

A placeholder that reads "the model behaves sensibly" is not one. That exact placeholder, on the refund action, cost SkyWays $2,000 on day 82.

Step 4 · Record the gate decision with its evidence#

A gate is a decision with evidence in front of a named person and their name on it, not a click, a status column or a meeting that ends in "fine". The record says what question the gate asks, what evidence was in the room, with denominators, and what would have made the answer no.

Step 5 · If you must cross on credit, write the waiver#

Some programmes genuinely cannot wait. Crossing the hard gate without its full set is allowed only with a written waiver: what is missing, why you are crossing now, the blast radius while it stands, what you are doing instead, and the date it expires. An undeclared crossing is indistinguishable from a completed one six weeks later.

Where you'll use it#

  • At the end of every P1, however small the change: the three artefacts may be one line each.
  • In the weekly review, re-asking the four questions of every soft decision: some turn hard as the build learns more.
  • When somebody proposes treating everything as hard "to be safe": that is how a build waits three weeks for a reversible framework choice.

Why it matters#

The hard gate is the difference between a limit that was decided and a limit that was enforced. SkyWays crossed P1 → P2 with all three artefacts written and the authority budget only partly implemented: the $400 cap was in the autonomy record and the prompt, and not in the refund tool. One column in the hand-off record (implemented, or written down?) would have caught it.

Try it#

A team has five open decisions at the end of P1: the model family; the refund approval threshold; the retrieval design; the dashboard layout; and whether a human approves cross-partner rebookings. Which are hard?

Show the answer

Two: the refund approval threshold and the approval rule for cross-partner rebookings. Both are autonomy decisions on consequential actions (money, and a rebooking that is hard to reverse), so everything downstream, from the tool signatures to the tests, depends on them, and they fail the first question. The model family can sit behind a gateway, the retrieval design behind a stub, and the dashboard can wait for P3; each needs only an owner and a date.

Key takeaways#

  1. Only P1 → P2 is hard: the spec, the bar per slice and the authority budget must be signed before P2 opens.
  2. Four questions, one "no" makes it hard: reversibility, a placeholder, an owner and a date, and downstream survival.
  3. Every soft decision needs a real placeholder, and every crossing on credit needs a written waiver.

FAQ#

What is the difference between a hard gate and a soft gate?#

A hard gate halts the work until its evidence exists; a soft gate lets work continue behind a placeholder while the decision is measured, provided it has a named owner and a date. In the agentic PDLC only the P1 → P2 hand-off is hard, along with the handful of decisions that fail at least one of the four questions.

What are one-way and two-way door decisions?#

The terms come from Amazon: a two-way door decision can be reversed cheaply, so it should be made quickly by a small group; a one-way door is hard or impossible to reverse, so it deserves care and evidence. The hard gate is a one-way door; the soft gates are two-way.

Doesn't a hard gate slow delivery down?#

One hard gate is fast; eleven are slow. Holding only the decisions that are genuinely expensive to reverse, and letting the rest run behind placeholders, is what lets the build start early without building on decisions that will move.

Who signs the hard gate?#

The solution architect is accountable for P1 and signs its exit. The product manager and architect jointly own the plan gate, the second of the five governance gates, whose evidence is the bolt cut, the authority budget and the gate map. A waiver for crossing without the full set is approved by the sponsor.

Apply it in your role#

If you are…Do thisThe AI-augmented shortcut
A forward-deployed engineerThe hard gate is where you protect the engagement: refuse to build past an unsigned autonomy decision, and put every soft one behind a placeholder with an owner.Have a model sort the open-decisions list with the four questions and draft the placeholder for each soft one.
A product manager or FDPMChair the gate: three hard decisions signed, every soft one with an owner and a date, any waiver in writing.Ask a model to produce the gate record from the decision log and flag decisions with no owner.
A GenAI or agentic AI engineerBuild behind an interface for every soft decision (framework, model, retrieval) so changing the answer is configuration, not a rewrite.Ask a coding agent for the interface and a stub for each soft decision, with a test that swaps the implementation.

Across the enterprise. Make P1 → P2 the one gate the portfolio enforces centrally, and let everything else run in parallel. One hard gate is how governance stops slowing delivery down.

The ten-minute workflow. Sort the open decisions before the gate meeting:

Here are our open decisions: <list>. For each, answer four questions: can it be reversed cheaply once
building starts? can the build proceed behind a placeholder? is there a named owner and a date? does
everything downstream survive if the answer changes? One "no" makes it HARD. Output a table
(decision, the four answers, HARD or SOFT) and the placeholder for each SOFT one.

Sources and credits#

IdeaOriginSource
One hard hand-off, three soft ones, and the four classifying questionsOriginal: this manualGates and Governance
The hard-gate waiverOriginal: this manualThe Agentic PDLC
Gates opened by evidenceBorrowedCooper, R. G. (1990). Stage-gate systems. Business Horizons 33(3)
One-way and two-way doorsBorrowedBezos, J. 2015 letter to Amazon shareholders. Type 1 and Type 2 decisions
Architecture decisions recorded where they are sensitiveBorrowedNygard, M. (2011). Documenting architecture decisions
The SkyWays figuresIllustrative: a fictional airlineThe workbench

Prefer GitHub? This track is on the wiki, and this lesson is markdown. Found a mistake? Edit this lesson, or say so.