Deciding Who Decides

Deciding Who Decides

Overview

Authority should sit where the evidence and competence are, not automatically at the top. But that gradient is bounded: an agent may decide reversible, well-evidenced work at the edge, and must escalate irreversible, trust-bearing, or thinly evidenced decisions to a human. This skill names, for a specific action, who decides and what makes it escalate. The point is to push decisions to the information, not to remove a human gate.

Decision contract

• Claim checked: the action's reversibility, evidence grade, and consequence place it at the agent's edge or behind a named human gate, with an escalation trigger concrete enough to obey and no required approval skipped.
• Artifact observed: the proposed action and target, the evidence and its grade, the consequence, and the agent's authority and standing gates -> a decision-rights line (action, who decides, escalation trigger), the evidence the decider must hold, and any mandatory human gate.
• Decision affected: block -- may the agent act at the edge or must it escalate: the decision-rights line and its escalation trigger.
• Failure class: misplaced-authority (an irreversible or thinly-evidenced action placed at the edge, or a gate skipped).
• Next action: escalate to the named human when consequence is protected and evidence is short of proven; stop when only the agent's assurance authorizes it.

When to Use

• An agent is about to act and it is unclear whether it may decide alone or must ask first.
• A change is irreversible, touches users, data, credentials, dependencies, agent authority, or a release.
• The evidence behind a decision is thin, contested, or rests only on the agent's own confidence.
• You are setting up an agent's standing authority and want explicit decision rights and escalation thresholds.

When Not to Use

• The edit is trivial and reversible with obvious proof and no new trust boundary.
• A required human approval already exists; this skill never exists to talk past it.
• An incident is live and you must stabilize first (use responding-to-incidents).

Inputs

• The proposed action, its target, and whether it can be undone.
• The evidence on hand, and how good it is (fact, local proof, source claim, or just confidence).
• The consequence if it is wrong, and who is affected.
• The agent's granted authority, and any standing rule or human gate that already applies.

Process

1. State the decision in one sentence and whether it is reversible.
2. Rate the evidence: proven, partially proven, or only asserted.
3. Rate the consequence: low, meaningful, or protected/irreversible.
4. Place the decision: reversible + well-evidenced + low consequence may be decided at the edge; anything irreversible, trust-bearing, or thinly evidenced escalates to a named human.
5. Name the escalation trigger in concrete terms an agent can actually obey.
6. Check that the placement raises rigor at the boundary; if it would let an agent skip a required gate, reject it.
7. Record who decides, who is informed, and what evidence the decider needs in hand. Name who is affected by the action and anyone who must be consulted before it — placement is hollow if a materially affected party never appears.

Outputs

• A decision-rights line: the action, who decides, and the escalation trigger.
• The evidence the decider must have before acting.
• Any human gate that stays mandatory regardless of the gradient.

Verification

• Every escalation trigger is concrete enough for an agent or reviewer to obey without judgment calls.
• No placement removes a required human approval or lets confidence stand in for evidence.
• The decider named actually has the evidence and authority the decision needs.

Escalation

• Escalate when consequence is protected or irreversible and the evidence is anything short of proven.
• Stop when the only thing authorizing the action is the agent's own assurance.
• Get an independent human when one agent's read is the sole basis for a trust-bearing decision.

Common Rationalizations

• "The agent has the most context, so it should decide." Local context is not judgment on an irreversible action.
• "It is faster if it just acts." Speed at the edge is the goal only while the work is reversible.
• "We delegated this already." Delegation sets a boundary; it does not dissolve the gate above the boundary.
• "It is probably fine." "Probably" is the trigger to escalate, not to proceed.

Red Flags

• Authority is argued from how confident or fluent the agent sounds.
• "Push authority to the information" is cited as a reason to skip a human approval.
• The escalation trigger is vague ("if it seems risky") rather than a named condition.
• An irreversible action is placed at the edge because the diff looked small.

Prompt

Decide who decides for this action the Nuclear-grade way.

Inputs:
- action and target:
- reversible? (yes/no):
- evidence and how good it is:
- consequence if wrong:
- agent authority / existing human gates:

Return:
- the decision in one sentence, and whether it is reversible
- evidence rating (proven / partial / asserted) and consequence rating (low / meaningful / protected)
- placement: who decides (agent at the edge, or a named human gate)
- the concrete escalation trigger an agent can obey
- any human approval that stays mandatory regardless of the gradient
- a check that the placement raises rigor at the boundary, not lowers it

Do not let confidence stand in for evidence. Do not use "authority to information" to skip a required gate.

Source-lineage note

This skill is an original software-workflow translation of decentralized-decision and decision-rights ideas. Pushing authority to where the information is takes concept inspiration from intent-based leadership (paraphrased, not template lineage; see docs/00-standards-foundation/do-not-cite-directly.md) and from public naval mission-command doctrine, bounded by the conservative-decision-making and questioning-attitude habits in DOE-HDBK-1028-2009, mapped in docs/00-standards-foundation/source-map.md. It does not create DOE compliance, formal assurance, safety, security, certification, or regulatory adequacy.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: FlyFission
• Source: FlyFission/nuclear-grade-context-engineering
• License: MIT

Install and usage instructions live in the source repository linked above.