LensReading which lens this session carries.

Improve · item 28

Agent lifecycle

An autonomy ceiling says how far an agent is allowed to go. It says nothing about how an agent gets there, or what pulls it back. This is the ladder between the two: five stages, each entered only against evidence named in advance, and a return path that fires on a threshold rather than waiting for someone to call a review.

Evaluation harness

Health

Where every managed agent sits

Agents by autonomy stage and lifecycle state, including the ones frozen, held or at a ceiling.

Scope of the program

The pilot covers Revenue Operations and Information Technology end to end. Those two were onboarded because their leaders volunteered, not because they carry the most risk. Human Resources and Legal — where a wrong autonomy call has the worst consequences — are outside the program entirely.

178 of 1,361 agents in the estate sit under lifecycle control — 13.1%. The other 1,183 have a published ceiling and nothing that governs movement toward it. The gate went live on Mar 3, 2026, so nothing that happened before that date has a transition record.

Under lifecycle control
178
of 1,361 agents · 1,183 outside
Transitions on record
227
206 up · 13 down · 8 refused
Promotions with evidence
206 / 206
every promotion cites the evaluation run that justified it
Demotions fired by rule
13 / 13
no return to a lower stage waited on a person to call it
Frozen at the gate
63
predate the gate, hold no entry evidence, cannot be promoted
Demotion rule not wired
29
above shadow with a written rule that nothing enforces

The five stages and what each one costs to enter

Entry evidence is named in advance and checked at the gate. Oversight is what the stage obliges while an agent sits in it. The return rule is what pulls an agent back without anyone deciding to.

1. Propose0agents hereHuman in the loop: Every output

The agent exists on paper and in code but touches no live work. It can read the estate and draft a recommendation into a scratch queue that nobody downstream consumes.

Entry evidence
  • ·Agent specification with a named purpose and a named owner
  • ·Risk classification recorded in the agent risk register
  • ·Data access scope reviewed and narrowed to what the purpose needs
Oversight while here

A human reads every output because nothing else does. There is no service level and no queue depth to defend.

Automatic return

There is nothing below propose. An agent that fails here is withdrawn rather than demoted, and the withdrawal is recorded as a transition so the attempt is not erased.

2. Shadow0agents hereHuman in the loop: Every output, compared

The agent runs against live volumes in parallel with the existing path and writes nothing. Its answer is recorded next to the answer the current process produced, and the two are compared.

Entry evidence
  • ·Shadow run log covering at least 500 live units
  • ·No-write attestation from the platform, signed by the platform owner
  • ·Agreement rate against the incumbent path, by unit type
Oversight while here

The incumbent path remains authoritative. Disagreements are triaged weekly by the process owner and become the first golden cases.

Automatic return

Agreement below 80 percent over a rolling 14 days returns the agent to propose and voids the shadow log, because a shadow that disagrees with reality is not evidence of anything except a specification problem.

3. Supervised55agents here, 17 of them frozenHuman in the loop: Every output, before release

The agent produces the working answer and a human releases it. The agent is doing the work; the human is the gate, not the reviewer of a sample.

Entry evidence
  • ·Golden test set defined and the first evaluation run recorded
  • ·Override capture wired so every human correction is stored as a case
  • ·Human queue capacity confirmed against forecast volume
  • ·Rollback path written down and owned
Oversight while here

A named human approves each release and every override is captured as a labeled example, which is where most of the golden set comes from.

Automatic return

Override rate above 25 percent over 14 days, or a failed evaluation run against the held-out set, returns the agent to shadow automatically.

4. Delegated50agents here, 15 of them frozenHuman in the loop: By exception and by band

The agent releases its own work inside an authority band. Anything above the band, and anything the agent flags as uncertain, routes to a human. Sampling replaces gating.

Entry evidence
  • ·Evaluation at or above the promotion floor on a held-out set the agent has never seen
  • ·Adversarial test executed against this agent with no unmitigated high finding
  • ·Rollback rehearsed on the line, with a measured time
  • ·Blast-radius cap set for any change affecting this agent
  • ·Service level attached with a named counterparty
Oversight while here

A sampling plan with a stated sample rate, an exception queue with an owner, and an override log that is read rather than merely written.

Automatic return

Two priority-one exceptions in a class this agent owns within 30 days, or an evaluation score below the floor, returns it to supervised automatically and notifies the function leader.

5. Autonomous73agents here, 31 of them frozenHuman in the loop: On exception only

The agent runs the line. Humans see exceptions, samples and the periodic evaluation, and nothing else. This is the ceiling for anything that is not people-affecting and does not sit inside a statutory decision.

Entry evidence
  • ·Independent evaluation review by someone outside the team that built the agent
  • ·Ninety days at delegated with the full override record intact
  • ·Kill switch tested on this agent, with a date and a measured stop time
  • ·Named owner for every exception class this agent can raise
  • ·Executive sign-off recorded against the accountable function leader
Oversight while here

Periodic re-evaluation on a fixed calendar, an exception taxonomy with named owners, and a kill switch that has actually been pulled in a drill rather than merely wired.

Automatic return

Any single priority-one exception attributed to the agent, an evaluation regression beyond the tolerance band, or a service-level breach attributed twice in a quarter returns it to delegated automatically.

Frozen at the stage they were found in

These agents were already operating when the gate went live. They were not demoted, because demoting a working agent to satisfy a process would have stopped real work. They are pinned instead: they keep their current stage and can never move up until a golden set exists for them.

63 frozen
AgentSub-functionPinned atSince
Password & Access AgentService DeskAutonomous7 months ago
Satisfaction AgentService DeskAutonomous10 months ago
Hierarchy Resolution AgentCRM Data StewardshipAutonomous10 months ago
Field Completeness AgentCRM Data StewardshipAutonomous4 months ago
Mover Adjustment AgentIdentity & AccessAutonomous3 months ago
Leaver Deprovisioning AgentIdentity & AccessAutonomous3 months ago
Identity & Access Challenger AgentIdentity & AccessSupervised4 months ago
Pricing OrchestratorPricing & Rate CardsSupervised10 months ago
Elasticity Modeling AgentPricing & Rate CardsSupervised3 months ago
Currency & Region AgentPricing & Rate CardsSupervised11 months ago
Pricing & Rate Cards Challenger AgentPricing & Rate CardsSupervised4 months ago
Software Deployment AgentEnd-User ComputeAutonomous10 months ago
Quote OrchestratorQuote-to-OrderDelegated10 months ago
Configuration Validation AgentQuote-to-OrderAutonomous3 months ago
Document Generation AgentQuote-to-OrderAutonomous9 months ago
Quote Hygiene AgentQuote-to-OrderAutonomous10 months ago
Order Management OrchestratorBooking & Order ManagementDelegated7 months ago
Credit Check AgentBooking & Order ManagementAutonomous3 months ago
Backlog AgentBooking & Order ManagementAutonomous10 months ago
Order Policy AgentBooking & Order ManagementSupervised4 months ago
Cost Optimization AgentInfrastructure & CloudAutonomous10 months ago
Configuration Drift AgentInfrastructure & CloudAutonomous4 months ago
Backup Verification AgentInfrastructure & CloudAutonomous4 months ago
Resilience Testing AgentInfrastructure & CloudAutonomous7 months ago
Billing Schedule AgentBilling InterlockAutonomous4 months ago
Usage Rating AgentBilling InterlockAutonomous7 months ago
Dispute Prevention AgentBilling InterlockAutonomous6 months ago
Proration AgentBilling InterlockAutonomous10 months ago
Topology Monitoring AgentNetworkDelegated3 months ago
Circuit Management AgentNetworkDelegated3 months ago
Release Coordination AgentApplication ManagementDelegated10 months ago
Lifecycle AgentApplication ManagementDelegated4 months ago
Contract Modification AgentRevenue Recognition InterlockSupervised6 months ago
Recognition Policy AgentRevenue Recognition InterlockSupervised7 months ago
Revenue Recognition Interlock Challenger AgentRevenue Recognition InterlockSupervised4 months ago
Auto-renewal AgentRenewals & Churn OpsAutonomous10 months ago
Churn Signal AgentRenewals & Churn OpsAutonomous10 months ago
Downgrade Processing AgentRenewals & Churn OpsAutonomous10 months ago
Renewals & Churn Ops Challenger AgentRenewals & Churn OpsSupervised3 months ago
Containment AgentCybersecurity OperationsDelegated7 months ago
Showing 40 of 63. The remainder carry the same condition.

The transition record

Every move up, every move down and every refusal, with who decided it, what evidence they held and whether a rule or a person made the call.

227 on record
LCT-0052RefusedSupervisedDelegatedAug 12, 2026
Device Policy Agent

Promotion requested and refused. The evidence pack was incomplete and the gate does not accept a promise to produce it later.

Clara Vogt · Quality Lead, Agent EvaluationEvidence: EVR-0112
LCT-0175RefusedSupervisedDelegatedAug 10, 2026
Forecast Integrity Challenger Agent

Promotion deferred by the owner pending an independent review that has not been scheduled.

Clara Vogt · Quality Lead, Agent EvaluationEvidence: EVR-0356
LCT-0141DemotionAutonomousDelegatedAug 9, 2026
Vulnerability Agent

An adversarial finding at high severity passed 30 days unmitigated. The agent was returned a stage until the mitigation is verified rather than merely deployed.

Fired by rule, no human decisionNo evaluation run attached
LCT-0197RefusedSupervisedDelegatedAug 7, 2026
Systems Policy Agent

Promotion requested and refused. The evaluation cited was run against the training set rather than the held-out set, which is not evidence of generalization.

Clara Vogt · Quality Lead, Agent EvaluationEvidence: EVR-0396
LCT-0024DemotionAutonomousDelegatedAug 6, 2026
Identity Orchestrator

An adversarial finding at high severity passed 30 days unmitigated. The agent was returned a stage until the mitigation is verified rather than merely deployed.

Fired by rule, no human decisionNo evaluation run attached
LCT-0079RefusedDelegatedAutonomousAug 5, 2026
Billing Interlock Orchestrator

Promotion requested and refused. The evaluation cited was run against the training set rather than the held-out set, which is not evidence of generalization.

Clara Vogt · Quality Lead, Agent EvaluationEvidence: EVR-0168
LCT-0111DemotionDelegatedSupervisedAug 3, 2026
Performance Obligation Agent

Two priority-one exceptions in a class this agent owns landed inside 30 days. The rule fired without a discretionary review, which is the point of writing it as a rule.

Fired by rule, no human decisionNo evaluation run attached
LCT-0146DemotionAutonomousDelegatedJul 31, 2026
Crediting Rules Agent

An adversarial finding at high severity passed 30 days unmitigated. The agent was returned a stage until the mitigation is verified rather than merely deployed.

Fired by rule, no human decisionNo evaluation run attached
LCT-0222DemotionAutonomousDelegatedJul 28, 2026
Revenue Analytics Orchestrator

A service-level breach was attributed to this agent for the second time in the quarter, which is one of the three conditions that return an autonomous agent to delegated without a review.

Fired by rule, no human decisionNo evaluation run attached
LCT-0021RefusedSupervisedDelegatedJul 27, 2026
CRM Data Stewardship Challenger Agent

Promotion requested and refused. The evidence pack was incomplete and the gate does not accept a promise to produce it later.

Clara Vogt · Quality Lead, Agent EvaluationEvidence: EVR-0048
LCT-0172DemotionDelegatedSupervisedJul 25, 2026
Integrity Policy Agent

Two priority-one exceptions in a class this agent owns landed inside 30 days. The rule fired without a discretionary review, which is the point of writing it as a rule.

Fired by rule, no human decisionNo evaluation run attached
LCT-0129DemotionAutonomousDelegatedJul 23, 2026
Security Operations Orchestrator

An adversarial finding at high severity passed 30 days unmitigated. The agent was returned a stage until the mitigation is verified rather than merely deployed.

Fired by rule, no human decisionNo evaluation run attached
LCT-0120RefusedDelegatedAutonomousJul 21, 2026
Renewal Ops Orchestrator

Promotion deferred by the owner pending an independent review that has not been scheduled.

Clara Vogt · Quality Lead, Agent EvaluationEvidence: EVR-0252
LCT-0161DemotionDelegatedSupervisedJul 20, 2026
Access Governance Agent

The scheduled evaluation scored below the floor after a model version change that had not been canaried. The agent dropped a stage the same night.

Fired by rule, no human decisionNo evaluation run attached

Held at the gate

Agents whose owner has asked for the next stage and been refused, with the specific artifact that is missing. A hold is a recorded decision, not an absence of one.

26 held · 13 ready
Service Desk OrchestratorDelegatedAutonomous

Held at this stage. The service level attached to this line has no counterparty signature, so there is nothing to breach and nothing to defend.

Kill switch tested on this agent, with a date and a measured stop timeNamed owner for every exception class this agent can raiseExecutive sign-off recorded against the accountable function leader
Data Stewardship OrchestratorDelegatedAutonomous

Held at this stage. The kill switch is wired but has never been pulled in a drill, so the stop time is asserted rather than measured.

Named owner for every exception class this agent can raiseExecutive sign-off recorded against the accountable function leader
Data Policy AgentSupervisedDelegated

Held at this stage. The service level attached to this line has no counterparty signature, so there is nothing to breach and nothing to defend.

Evaluation at or above the promotion floor on a held-out set the agent has never seenAdversarial test executed against this agent with no unmitigated high findingRollback rehearsed on the line, with a measured timeBlast-radius cap set for any change affecting this agentService level attached with a named counterparty

Held at this stage. No adversarial test has been executed against this agent, and delegated entry requires one with no unmitigated high finding.

Rollback rehearsed on the line, with a measured timeBlast-radius cap set for any change affecting this agentService level attached with a named counterparty
Duty Conflict AgentSupervisedDelegated

Held at this stage. Override capture is not wired on this line, so the corrections humans make are lost rather than becoming labeled cases.

Service level attached with a named counterparty
Device Policy AgentSupervisedDelegated

Held at this stage. No adversarial test has been executed against this agent, and delegated entry requires one with no unmitigated high finding.

Rollback rehearsed on the line, with a measured timeBlast-radius cap set for any change affecting this agentService level attached with a named counterparty

Held at this stage. No adversarial test has been executed against this agent, and delegated entry requires one with no unmitigated high finding.

Adversarial test executed against this agent with no unmitigated high findingRollback rehearsed on the line, with a measured timeBlast-radius cap set for any change affecting this agentService level attached with a named counterparty
Infrastructure OrchestratorDelegatedAutonomous

Held at this stage. Rollback has been written down for this line but never rehearsed, so the recorded recovery time is an estimate rather than a measurement.

Kill switch tested on this agent, with a date and a measured stop timeNamed owner for every exception class this agent can raiseExecutive sign-off recorded against the accountable function leader
Infrastructure Policy AgentSupervisedDelegated

Held at this stage. The service level attached to this line has no counterparty signature, so there is nothing to breach and nothing to defend.

Blast-radius cap set for any change affecting this agentService level attached with a named counterparty
Billing Interlock OrchestratorDelegatedAutonomous

Held at this stage. The held-out evaluation set has never been run against this version, so there is no evidence to attach to a promotion record.

Independent evaluation review by someone outside the team that built the agentNinety days at delegated with the full override record intactKill switch tested on this agent, with a date and a measured stop timeNamed owner for every exception class this agent can raiseExecutive sign-off recorded against the accountable function leader

Held at this stage. The kill switch is wired but has never been pulled in a drill, so the stop time is asserted rather than measured.

Blast-radius cap set for any change affecting this agentService level attached with a named counterparty
Network Policy AgentSupervisedDelegated

Held at this stage. The held-out evaluation set has never been run against this version, so there is no evidence to attach to a promotion record.

Blast-radius cap set for any change affecting this agentService level attached with a named counterparty
Network Challenger AgentSupervisedDelegated

Held at this stage. The independent evaluation review is outstanding — the only review on file was performed by the team that built the agent.

Evaluation at or above the promotion floor on a held-out set the agent has never seenAdversarial test executed against this agent with no unmitigated high findingRollback rehearsed on the line, with a measured timeBlast-radius cap set for any change affecting this agentService level attached with a named counterparty

Held at this stage. The independent evaluation review is outstanding — the only review on file was performed by the team that built the agent.

Evaluation at or above the promotion floor on a held-out set the agent has never seenAdversarial test executed against this agent with no unmitigated high findingRollback rehearsed on the line, with a measured timeBlast-radius cap set for any change affecting this agentService level attached with a named counterparty
Renewal Ops OrchestratorDelegatedAutonomous

Held at this stage. Rollback has been written down for this line but never rehearsed, so the recorded recovery time is an estimate rather than a measurement.

Kill switch tested on this agent, with a date and a measured stop timeNamed owner for every exception class this agent can raiseExecutive sign-off recorded against the accountable function leader
Comp Ops Policy AgentSupervisedDelegated

Held at this stage. Override capture is not wired on this line, so the corrections humans make are lost rather than becoming labeled cases.

Rollback rehearsed on the line, with a measured timeBlast-radius cap set for any change affecting this agentService level attached with a named counterparty
Showing 16 of 26 held.

Where the program does and does not reach

Two functions are onboarded end to end. Twelve are not. Reading the estate as governed because two functions are governed is the mistake this panel exists to prevent.

Revenue Operations88 agents
2 evidence complete14 held37 at ceiling35 frozen79 of 88 have the demotion rule wired
Information Technology90 agents
11 evidence complete12 held39 at ceiling28 frozen70 of 90 have the demotion rule wired
Outside the program

1,183 agents across the other twelve functions have a published autonomy ceiling and no lifecycle record at all. Human Resources and Legal — the two functions where a wrong autonomy call does the most damage — are not in the pilot. Nothing on this page should be read as covering them.

Actions

What is waiting on a person

Promotion is the only way autonomy increases on this estate, and no agent can promote itself.

Operations

What this desk is allowed to start

A surface that only reports is not operable. This is the work this page can set in motion, and the bound it runs into.

Trigger and bound

This desk can run the evidence checks a promotion requires, arm a demotion trigger, and record a stage change once an approver signs it. It cannot promote, demote, or lift a ceiling on its own — every stage change on this estate carries a human name.

Live observability

What the record shows right now

Current autonomy stage of every managed agent.

Current distribution

178 agents

Autonomous7341%
Delegated5028%
Supervised5531%

Is policy and strategy coming to fruition

Whether the written intent is holding here

73 of 178 managed agents run autonomously. 13 are ready to promote and waiting.

Not holding on the record

The strategy position is that autonomy is earned against evidence rather than granted by default. That holds structurally: 41% run autonomously, and every one of them passed a recorded gate to get there. Where it bites is the other direction — 76 agents are at a ceiling their class will not let them pass, 63 are frozen and 26 are held, and each of those is a human decision that has not been revisited.

What this page is, and what it is not

Every evaluation on this page ran against golden sets we wrote ourselves, on a modeled estate. A passing suite proves an agent behaves the way we specified, not that the specification is right, and no evaluation here has been reviewed by anyone outside the team that built the agent.

The narrow claim this page can defend is this: 206 of 206 promotions on record cite an evaluation run that cleared the floor, and no promotion was granted against a run that did not. It cannot claim the estate is governed, because 1,183 agents sit outside the program. Nor can it claim every agent inside the pilot earned its stage: 63 were already running when the gate went live and hold no entry evidence at all, and a further 12 hold a golden set that has never once cleared the floor, so there is no result their stage could cite.

The demotion rule is the weakest part and is shown as weak rather than green. 149 of 178 agents have it wired to a live threshold that fires on its own; 29 sit above shadow with a rule that is written down and enforced by nobody. 13 of 13 demotions on record fired on a threshold rather than on a person noticing. A rule that needs a person to notice is a review, not a rule.