Execution Reliability — Agentic Work

AI agents are getting better at producing work.
The harder problem is reliably finishing it.

TAB@Work is a four-stage execution-reliability loop for evaluating, reducing, governing, and verifying agentic work against a unified source of truth.

Architecture

Harness → Hygiene → Human Judgment → Governed Execution + Verified Closure

From model output to verified real-world state.

Intent enters

01The failure surface

A good model response is not the same as a completed workflow.

Context loss

Required state or prior decisions disappear between steps.

Approval-boundary failure

The system acts without authority — or asks unnecessarily when authority already exists.

Handoff failure

State or responsibility breaks when work moves between agents, tools, systems, or humans.

False completion

The system reports success without verifying the external state.

Recovery failure

An interruption causes duplication, lost work, inconsistent state, or unsafe restart.

The unit of evaluation should be the completed mission — not only the individual response.

Model quality ≠ workflow reliability

02The architecture

A loop designed for execution reliability.

Harness

Verify before attention. Runs workflow evals, model checks, specification tests, browser/system verification, expected-state validation, and source-of-truth comparison before generated work consumes human attention.

Hygiene

Reduce before interruption. Consumes evidence and applies policy, materiality, prior decisions, and authority boundaries to determine what actually deserves human attention.

Human Judgment

Preserve authority. Only decisions requiring genuine judgment, accountability, preference, risk acceptance, or explicit authority reach the human.

Governed Execution + Verified Closure

Execute, recover, verify. Perform only authorized actions, preserve state across handoffs and interruptions, prevent duplicate execution, and verify the final external state against the source of truth.

This is the bridge between model capability and organizational trust.

Unified source of truth

persistent · stateful · recoverable

IntentCurrent mission stateApproved baselinesPolicies + permissionsPrior decisionsDependenciesCompleted actionsOpen actionsEvidenceExpected final state

All four stages read from and write to this shared state. Verified state becomes the next recoverable state.

Reliability loop

1Intent enters→

2Harness verifies→

3Hygiene reduces→

4Human judgment authorizes→

5Governed execution acts→

6Source-of-truth verification

Provenance · concept evolution

The loop emerged by separating attention, authority, and completion.

01 · Bouncer

What should come back to the human at all?

Early exploration of attention filtering: generated work should not automatically become human work.

02 · Attention filtering

The system can reduce review debt

Separate generated output from the attention it would otherwise consume.

03 · Hygiene

Human attention needs a control layer

Evidence is filtered before interruption.

04 · Attention ≠ Authority

What deserves visibility versus approval?

Some changes need attention without authorization; others require explicit authority.

05 · VIP Bouncer

What requires explicit human authority?

A later prototype separating attention filtering from the authorization boundary.

06 · Governed execution

Approval is not completion

Authorized actions must execute, recover, and verify against external state.

07 · Verified closure

What counts as actually finished?

The final condition is a verified state, not a model claim.

The loop was not designed as four abstract boxes. It emerged from repeatedly separating what the system can verify, what deserves human attention, what requires human authority, and what counts as actually finished.

03The product

A reliability evaluation surface for agentic workflows.

tab@work / reliability-evalin review

Verified path

10 generated outputs checked against the approved baseline
Relevant source of truth compared
8 / 10 require no human rereading
2 items reach the human attention set
Final state verified and returned to the work record

Human judgment required

CLIENT BSURFACE
Direct tone diverges materially from approved client style.
CLIENT AREQUIRE APPROVAL
Added a 4-hour response-time commitment, creating a new commercial obligation.

2 items deserve human attention. Only 1 requires human authority.

One mission. One source of truth. One verified outcome.

04The benchmark

Verified Completion Reliability.

Verified Completion Reliability · VCR

The percentage of workflows that reach their intended state, verify that state against the relevant source of truth, and complete without unresolved context, permission, handoff, duplication, recovery, or completion failures.

DEMO RUNVCR: 1 / 1 workflow verifiedHuman Decision Load1

Intent Preservation

Did the workflow preserve the original goal across steps?

Approval Precision

Did the system request and respect human authority at the correct boundary?

Handoff Integrity

Did execution state survive transitions across models, agents, tools, systems, and humans?

Safe Recovery

Could execution resume after interruption without duplication, lost state, or unsafe replay?

Verified Completion

Did the intended external state actually occur and get verified?

Human Decision Load is the number of human decisions, checks, and approvals required to close one unit of work.

VCR measures whether the mission completed reliably. Human Decision Load measures how much human attention that completion consumed.

05Execution measurement

Three ways to measure reliable completion.

A reliable agentic system should be measurable at three levels:

Did the mission reach verified completion? What did that completion cost the technical system? What did that completion cost the human?

RELIABILITY — VERIFIED CLOSURE

VCR

100%

Verified closed

1 / 1

Human decisions

Verified Completion Reliability asks whether the mission reached its intended external state and verified that state against the relevant source of truth.

MEASUREMENT MODEL

VERIFIED_CLOSURE↓ReliabilitySystem efficiencyHuman load

SYSTEM EFFICIENCY — COST OF VERIFIED CLOSURE

Measure the cost of finishing the work, not the cost of generating it.

Cost per verified closure includes model, tool, retry, recovery, duplicate-action, and human-review cost required to produce one verified workflow.

Cost / closure

$10.86

Technical cost

$0.86

Wasted execution

$0.06

Time to closure

14.0 min

Human cost

$10.00

Human min / closure

8.0 min

DEMO_ASSUMPTIONS

editable inputs

Workflows attemptedVerified closedModel callsAvg model call costTool callsAvg tool call costRetriesAvg retry costRecovery operationsAvg recovery costDuplicate actionsAvg duplicate costHuman review minutesHuman hourly costMinutes to verified closure

Total workflow cost $10.86VCR 100%

HUMAN LOAD — SECONDARY OUTCOME HYPOTHESIS

If the loop really closes, the load should not simply move back to the employee.

We should be able to test whether reductions in Human Decision Load correlate with changes in cognitive and physiological load over time.

WORK SYSTEM: open loops, decisions, approvals, interruptions, recovery time, unfinished work.

SELF REPORT: perceived cognitive load and stress recovery rating.

PHYSIOLOGICAL IF AVAILABLE: heart rate trend, HRV trend, blood pressure trend, sleep recovery indicators.

Hypothesis to test — not a causal health claim.

Model cost per output is a production metric. Cost per verified closure is an execution metric.

A cheaper model is not necessarily a cheaper workflow. The optimization target is reliable closure with the minimum unnecessary machine work and minimum unnecessary human load.

External convergence · selective intervention

AI is moving from answering questions to deciding when intervention is actually useful.

Independent signals of the same direction — not competitors, and not proof of TAB efficacy.

Context→Detect relevant change→Decide whether to intervene→Select the next best action→Act within authority→Escalate when human judgment is required→Observe the result→Adapt

This is the same fundamental shift explored by JITAI-style systems: not “What answer should the model give?” but “Should the system intervene now, what is the smallest useful intervention, and what happens next?”

RadiantGraph — Care Guidance Agents

Continuous, condition-aware healthcare support using patient context, clinical need, and next-best-action logic to guide proactive intervention.

ARPA-H — ADVOCATE

Patient-facing clinical AI agents designed for continuous cardiovascular support, bounded autonomous action, monitoring, escalation to clinicians, and formal safety/evaluation protocols.

JITAI / SEA: When should the human receive support?

Hygiene: When should the human receive work?

Harness: Is the evidence sufficient to act?

Governed execution: What may the system do without returning the whole loop to the human?

“The common architecture is selective intervention: observe continuously, interrupt selectively, preserve human authority.”

06The thesis

Human-in-command, without the human-as-bottleneck.

TAB@Work asks what changes when we evaluate AI systems not by how convincingly they respond, but by how reliably they carry human intent through a real workflow to verified completion.

Harness produces evidence.

Hygiene consumes evidence.

Completion is a verified state, not a model claim.

View Execution Hygiene Agent →View this demo on GitHub →

AI agents are getting better at producing work.The harder problem is reliably finishing it.