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 decision
An open decision
- Cheap to reverse once building starts?
- yes
- Can the build proceed behind a placeholder?
- yes
- A named owner and a date?
- yes
- Does downstream work survive a change?
- yes
- SOFTruns beside the build, behind a placeholder
- any noHARDsettle it before the phase closes
- P1 to P2: the spec, the bar and the guardrails
- Soft gates cross with a placeholder, an owner, a date
- 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.
| Artefact | Why nothing downstream survives without it | What happens if it is missing |
|---|---|---|
| The eight-field spec | It is what the coding agent reads, and a machine cannot ask what you meant | The 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 enough | The team ships at whatever number it got and calls that the bar |
| The authority budget and gate map | It lists what the machine may do, with caps typed into signatures | The 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.
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 decision | Why it halts |
|---|---|
| The AI-fit verdict | It fixes the shape of the product; changing it later means starting again |
| The autonomy level on money actions | It moves money, and a regulator has a rule about it |
| The spec, the bar and the guardrails | Everything 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#
- Only P1 → P2 is hard: the spec, the bar per slice and the authority budget must be signed before P2 opens.
- Four questions, one "no" makes it hard: reversibility, a placeholder, an owner and a date, and downstream survival.
- 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 this | The AI-augmented shortcut |
|---|---|---|
| A forward-deployed engineer | The 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 FDPM | Chair 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 engineer | Build 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#
| Idea | Origin | Source |
|---|---|---|
| One hard hand-off, three soft ones, and the four classifying questions | Original: this manual | Gates and Governance |
| The hard-gate waiver | Original: this manual | The Agentic PDLC |
| Gates opened by evidence | Borrowed | Cooper, R. G. (1990). Stage-gate systems. Business Horizons 33(3) |
| One-way and two-way doors | Borrowed | Bezos, J. 2015 letter to Amazon shareholders. Type 1 and Type 2 decisions |
| Architecture decisions recorded where they are sensitive | Borrowed | Nygard, M. (2011). Documenting architecture decisions |
| The SkyWays figures | Illustrative: a fictional airline | The workbench |