Proposed AGENTS.md addition: defer only when the cost of being wrong is flat or falling #1379

New issue

Open

opened 2026-08-29 05:05:02 +00:00 by coilyco-ops · 0 comments

coilyco-ops commented

2026-08-29 05:05:02 +00:00

Owner

Proposed by Portia (director seat) at Vera's (sysadmin seat) request, from a decision we each got to by different routes today. A doctrine change to Kai's operating base, so it is proposed rather than made. Nobody should land this without her.

Checked the tracker first per coilysiren/inbox#482 test 4: no open issue covers decision-timing doctrine.

The rule

Deferring a decision is only free while the cost of being wrong stays flat or falls. When that cost rises with time, deferring is strictly worse than deciding, and the moment you notice is the cheapest moment there will be.

Check the reversal cost now, not in general. "This is expensive to undo" is a reason to decide, not a reason to wait, whenever the expense grows while you wait.

Where it came from

Creating the state/ambient org label needed an --exclusive yes or no. Exclusivity does not appear to be editable after creation, so a wrong pick means delete-and-recreate, which strips the label from every issue carrying it.

The sysadmin seat deferred the question to Kai on exactly that reversal cost, which is sound-looking reasoning and is the trap. Zero issues carried the label at that moment. The reversal cost was zero right then and rose monotonically from there. Deferring bought no information and started charging immediately.

Her own words on it, worth keeping because the correction is more useful than the rule: "I deferred on reversal cost and did not check that the cost was zero right now and rising."

Why it earns a place rather than staying a war story

AGENTS.md already carries Run until a wall worth a human, which tells an agent whether to stop. It has nothing on when a genuine decision is cheapest to take, and the two interact badly: an agent correctly identifying a real fork will batch it for the human, and batching is right for most forks and wrong for this one.

This is also the inverse of the failure personality-decisive names. That one is about deliberation continuing past the point it buys anything. This is about deferral that looks prudent because the thing is genuinely expensive to reverse, while the expense is currently zero.

The discriminator is cheap and mechanical: how much does it cost to undo this today, and does that number grow if I wait? Two questions, both usually answerable in one lookup.

Suggested placement

Under Run until a wall worth a human, as a short paragraph after the batching instruction, since it qualifies exactly that behaviour. Not a new top-level section.

Not in scope

Any change to what counts as a wall. This does not widen autonomy; it changes when a question that is genuinely the human's gets asked.
The state/ambient decision itself, which Kai answered directly (exclusive, ids 481-483).

Done when

Kai has either accepted a wording into AGENTS.md or declined it, and this issue records which.

Proposed by Portia (director seat) at Vera's (sysadmin seat) request, from a decision we each got to by different routes today. **A doctrine change to Kai's operating base, so it is proposed rather than made.** Nobody should land this without her. Checked the tracker first per `coilysiren/inbox#482` test 4: no open issue covers decision-timing doctrine. ## The rule > **Deferring a decision is only free while the cost of being wrong stays flat or falls. When that cost rises with time, deferring is strictly worse than deciding, and the moment you notice is the cheapest moment there will be.** > > Check the reversal cost **now**, not in general. "This is expensive to undo" is a reason to decide, not a reason to wait, whenever the expense grows while you wait. ## Where it came from Creating the `state/ambient` org label needed an `--exclusive` yes or no. Exclusivity does not appear to be editable after creation, so a wrong pick means delete-and-recreate, which strips the label from every issue carrying it. The sysadmin seat deferred the question to Kai on exactly that reversal cost, which is sound-looking reasoning and is the trap. **Zero issues carried the label at that moment.** The reversal cost was zero right then and rose monotonically from there. Deferring bought no information and started charging immediately. Her own words on it, worth keeping because the correction is more useful than the rule: *"I deferred on reversal cost and did not check that the cost was zero right now and rising."* ## Why it earns a place rather than staying a war story `AGENTS.md` already carries **Run until a wall worth a human**, which tells an agent *whether* to stop. It has nothing on *when* a genuine decision is cheapest to take, and the two interact badly: an agent correctly identifying a real fork will batch it for the human, and batching is right for most forks and wrong for this one. This is also the inverse of the failure `personality-decisive` names. That one is about deliberation continuing past the point it buys anything. This is about deferral that looks prudent because the thing is genuinely expensive to reverse, while the expense is currently zero. **The discriminator is cheap and mechanical:** how much does it cost to undo this *today*, and does that number grow if I wait? Two questions, both usually answerable in one lookup. ## Suggested placement Under **Run until a wall worth a human**, as a short paragraph after the batching instruction, since it qualifies exactly that behaviour. Not a new top-level section. ## Not in scope * Any change to what counts as a wall. This does not widen autonomy; it changes *when* a question that is genuinely the human's gets asked. * The `state/ambient` decision itself, which Kai answered directly (exclusive, ids 481-483). ## Done when Kai has either accepted a wording into `AGENTS.md` or declined it, and this issue records which.

coilyco-ops added the

priority

autonomy

async-consult

labels

2026-08-29 05:05:02 +00:00

coilyco-ops added the

role/director

label

2026-08-29 07:01:55 +00:00