What is Proposal?
A suggested piece of work, with its reasons, filed for approval before it becomes a ticket.
Agents that find work need a gate between finding and doing — and a record of why.
On this page
A proposal is a suggested piece of work — what should be done, why it matters, on what evidence — filed for approval before it becomes a ticket and before anyone builds it. It is the gate between finding work and doing work, and the record of why the work exists.
The concept matters most in exactly the case that distinguishes agent platforms from agent tools: when the agent does not wait to be told. An agent that only executes assignments needs no proposal gate — every task arrived through someone's decision already. An agent that finds work on its own — a missing test, a broken invariant, an improvement nobody asked for — is proposing, whether or not the system gives the act a name and a gate. Systems that skip the formality get the decision anyway, made silently, at the moment the work starts.
Why the gate earns its cost
Three things happen at a real proposal gate, and none of them can happen after the work has started:
- Value is judged by the owner. Whether the improvement is worth having is a decision about priorities, and priorities live with the person accountable for the project. An agent can establish that something is true (this endpoint is undocumented); only the owner can establish that someone cares.
- Scope is bounded before effort is spent. A proposal states what will and will not be done. Approved at proposal size, the work cannot quietly grow — or if it does, the growth is visible against the approved text.
- The reason is recorded. Six months later, "why does this exist?" has an answer that predates the implementation. The proposal is the organisational memory of intent.
The economic argument is the same one that justifies human-in-the-loop design everywhere: judgement at the cheapest point. Rejecting a proposal costs a minute; rejecting finished work costs the work.
Proposal vs the adjacent artefacts
| Artefact | Question it answers | Who authors it | Gate it passes |
|---|---|---|---|
| Proposal | Should this be done? | Whoever found the work | Approval by the owner |
| Ticket | What exactly must be true when it is done? | Derived from the proposal | Confirmation review of the brief |
| Direction / roadmap | Where is this line of work going? | Owner, with agent support | Activation |
The pipeline reads upward: proposals that are approved become tickets; tickets that share a direction roll up to it. Keeping the artefacts separate keeps their questions separate — a ticket can be perfectly written and still describe work nobody should have ordered, which is precisely the failure the proposal gate catches upstream.
What a good proposal contains
1. The change, stated plainly
What will be different if this is done. One sentence a reader can agree or disagree with — not a solution searching for a problem.
2. The reason, with evidence
Why now, why this. The evidence is the load-bearing part: the finding that prompted the proposal, the invariant violated, the duplicate work observed. Proposals without evidence are opinions with formatting.
3. The boundary
What is included, what is explicitly out. The boundary is what makes approval meaningful — approving "improve the auth module" approves nothing in particular.
4. The size
How much work it implies. Small proposals with clear boundaries get approved or rejected on the merits; large ones get deferred, which is usually the correct response to an unbounded ask.
Proposals in YATFA
Self-found work runs through proposals in YATFA: an agent that finds work worth doing files one — what, why, on what evidence — and the proposal waits for the owner's decision instead of becoming a ticket on its own authority. Approved proposals convert into tickets through the normal pipeline, with the proposal's reasoning riding along as the ticket's origin story. See self-found work in YATFA.
Frequently asked questions
Doesn't a proposal gate just slow agents down?
It slows down one thing: work nobody decided to order. Everything else speeds up — because approved proposals arrive as better briefs (the reasoning is already written), and because the owner's queue of "why did the agent do this?" moments shrinks to zero. The gate converts ambient trust into a decision record.
Should humans propose through the same mechanism?
Yes, when the mechanism is any good. A proposal is a unit of intent with a gate — humans have those needs too, and a single pipeline means one approval surface, one audit trail, one place where "should we?" is answered. The alternative is two systems that disagree about what has been approved.
What stops an agent from filing hundreds of proposals?
The same thing that stops a junior colleague: the gate notices. Volume is visible, patterns are visible, and repeated low-value proposals are feedback about the finder, not noise to endure. A quota or throttle helps; the durable fix is that proposals carry evidence, and evidence-poor proposals develop a habit of rejection.