Heat is the consequence an action carries if it goes wrong. Five dimensions decide how hot. Heat decides who pulls the trigger. One ladder, every function — the internal rubric that turns "should the agent be allowed to do this?" from a debate into a reading.
Most teams argue about agent autonomy with vibes — "that feels risky." The ladder replaces vibes with five questions you can answer about any action, in any function. Score each from 0 (cool) to 4 (hot).
How hard is it to undo once done — a click, a cleanup job, or never?
0 · fully undoable 4 · permanent, unrecoverableOne record, one customer, one system — or everything at once?
0 · single item 4 · the whole estateDoes the output stay internal, reach one outsider, or go public?
0 · nobody outside the loop 4 · public, press, regulatorsDoes the action move money, make a promise, or sign the company's name?
0 · binds nothing 4 · legally / financially bindingRead-only, a scratchpad, an internal system — or production and access control?
0 · read-only 4 · production & permissionsNo averaging. Four cool dials never buy back one hot one — a perfectly reversible, narrow, internal action that signs a contract is still a contract. The hottest single dimension sets the band, and the band sets the control posture. This one rule is what stops "but it's mostly fine" from shipping a disaster.
Heat is only useful if it changes something. Each band maps to exactly one answer to the only question that matters: who pulls the trigger? Tap a rung to open it.
The same verb can live on three different rungs. Context is what moves it — these are the forces that push heat up, and the design moves that pull it back down. De-escalators are the real product: they're how you unlock autonomy safely instead of just forbidding things.
Each of these pushes one dimension toward red.
Each of these is a design decision that cools a dimension.
Every function runs the same spectrum — from cool drafting to critical, irreversible action. Filter by function to see how the ladder lands in finance, engineering, support, HR, and beyond. Tap any action to drop it into Heat Check below.
Notice the pattern: cool drafting and recommending appears in every column, and critical, irreversible action caps every column. The function changes; the shape of the ladder does not.
Type any agent action in plain language. Heat Check reads it across the five dimensions, places it on the ladder, and tells you the control posture it deserves. The guess isn't gospel — click any cell to correct it and watch the band move.
A heat model is only worth something when it changes what an agent is allowed to do unsupervised. Five steps, one working session, a living artifact.
Inventory every action your agents take or will take. Verbs and objects — "send quote", "update record" — not capabilities.
Run each action through the five dimensions. Disagreement is signal: it means the action is underspecified.
Hottest dimension sets the band. The band sets the posture. Write both next to every action — no exceptions, no averaging.
For every action stuck at HIGH or CRITICAL, ask which de-escalator would cool it — drafts, sandboxes, caps, undo windows.
The ladder is living policy. Every incident and every quarter, re-score. Promote actions that earned trust; demote ones that burned it.
auxfirst designs the trust layer for agentic systems — visible intent, evidence, autonomy boundaries, escalation paths, audit trails. We'll run your real action inventory through the ladder and hand you the rubric your team operates from.