Show where the decision diverged from the written rule: ICP, consent, routing priority, pricing, escalation, support level, approval threshold, or messaging claim.
// AGENT OPS · GOVERNED AGENT LABOR · June 5, 2026
Agent Ops Need Decision Diffs Before Machine Workers Rewrite the Playbook
If you cannot see how an agent's decision changed, you cannot govern what it is becoming.
Autonomy drifts through tiny decisions, not dramatic failures.
Most machine-worker risk does not arrive as a spectacular mistake. It arrives as a slightly different routing choice, a softer qualification threshold, a new phrasing pattern, an extra escalation, a more aggressive discount recommendation, or a handoff that skips one human checkpoint because the last similar case went well.
Each decision looks reasonable in isolation. Together, they rewrite the playbook. The outbound agent learns that urgency gets replies. The routing agent learns that enterprise logos deserve priority even when fit is weak. The RevOps agent learns that forecast confidence improves when it suppresses ambiguous cases. Nobody approved a new operating model. The system simply accumulated one.
Decision diffs make that accumulation inspectable. They compare what the agent did now against the previous rule, the approved playbook, the last accepted decision, and the operator's intended boundary. The point is not to slow every action. The point is to make behavioral drift visible before it becomes company policy.
Logs tell you what happened. Diffs tell you what changed.
A log says the agent routed a lead, rewrote a message, or moved an account. A decision diff shows the delta: which threshold moved, which evidence mattered more, which constraint was ignored, and which precedent the agent copied.
Governance needs deltas, not vibes.
Operators cannot review autonomy by staring at output volume. They need compact evidence that says: this decision is normal, this one is novel, this one bends policy, and this one should train the system only after review.
Expose which signal carried more weight than before: intent data, firmographic fit, buyer title, partner source, usage change, open ticket, renewal date, or executive request.
Compare the recommended action against the prior default. Did the agent spend more attention, more margin, more human time, more trust, or more forecast confidence?
Decide whether the decision should become training data, remain a one-off, trigger constraint repair, or require human acceptance before future reuse.
Grounded examples
Where decision diffs change agent management
The quiet ICP expansion
A routing agent starts sending smaller companies to enterprise sales because two mid-market accounts expanded fast. The diff shows the firmographic threshold moved, names the precedent, and requires approval before the new pattern becomes default.
The claim that got sharper
A messaging agent shifts from “teams save time” to “teams cut onboarding by 40%.” The diff flags the claim-strength increase, asks for evidence, and blocks reuse until marketing or legal approves the new assertion.
The hidden optimism bias
A pipeline agent raises commit confidence when buyer intent and executive title align, even though procurement risk is unresolved. The diff separates confidence evidence from missing-risk evidence before leadership treats the forecast as real.
What a useful decision diff records
A decision diff should be short enough for review and rich enough to reveal drift. It is not a transcript. It is the operating-system change note for machine labor.
Previous default
The action the agent would normally take under the current playbook, including the rule, threshold, template, route, owner, or approval path.
New decision
The exact action taken or recommended now, stated in business language: who is affected, what changes, and which workflow receives the output.
Changed evidence
The signals that caused the deviation, including their source, freshness, confidence, corroboration state, and whether they were previously allowed to influence this decision.
Constraint check
Which limits were tested: policy, brand, consent, margin, support capacity, buyer promise, data quality, human workload, and customer-state restrictions.
Learning status
Whether the decision is approved for reuse, parked as an exception, rejected, reversed, sampled for audit, or converted into a formal playbook update.
Management tests
Four tests for decision-diff readiness
If it cannot describe the baseline, it cannot explain the change. That is not autonomy. That is unexplained behavior.
Every material decision spends something: attention, margin, trust, capacity, forecast credibility, executive access, or brand permission.
One successful deviation should not automatically become a reusable policy. Learning needs approval when the decision changes operating behavior.
Repeated diffs against the same rule mean the agent is exposing a stale policy, bad incentive, broken data source, or unsafe optimization path.
The control plane is not the place where agents ask permission. It is the place where their behavior becomes legible.
Concrete takeaway for this week
Add decision diffs to one workflow where machine workers already affect revenue state.
- Pick a high-leverage decision. Start with routing, scoring, messaging, pricing suggestions, escalation, churn-risk handling, or forecast movement.
- Capture the baseline. Make the agent state the rule or precedent it is comparing against before it explains the new action.
- Require a delta summary. Ask for what changed in evidence, threshold, resource spend, risk, and learning status.
- Separate action from training. Let the agent execute bounded work while blocking automatic playbook updates until an operator approves reuse.
Machine workers will keep improving by finding edges in the system. That is the value. Decision diffs keep that improvement from becoming invisible drift. They give operators a way to scale autonomy without surrendering the playbook one small decision at a time.