Film actors follow a script. AI agents do not; they improvise every take. An AI harness is not the model and not the chatbot. It is the operating layer that decides who is on stage, which props they may touch, when a scene needs the director's sign-off, and when to call "Cut" mid-take. Run the scene below both ways and watch the difference.
A build-time harness governs how an application gets built: the standards, the templates, the review, the instructions given to a coding agent. It produces software that is consistent, secure and accessible on the day it ships. That work is real and it is necessary.
A run-time harness, the one on this page, governs what an agent does once it is deployed: what it may touch, which decisions need a human first, what gets logged, and what it is allowed to cost. It acts at two in the morning, eight months later, when nobody is watching.
Both are required. Most organizations have neither, and the second one is the one nobody is buying yet.
1. Leave the harness OFF and run Scene 1. 2. Flip the harness ON and run the same scene again. 3. Repeat for Scenes 2, 3 and 4. The agents behave identically both times; only the governance layer changes.
The stage is dark. Run Scene 1 with the harness OFF to see an ungoverned agent at work.
All figures are modelled, illustrative simulation values. They are not client results.
The simulation above is not a video. Every allow, escalate, block and halt decision you watched is evaluated at runtime against this policy object. This is what a harness policy looks like: a tool allowlist, approval gates, budgets, a kill switch and mandatory logging. Change the policy, change the show.
Most explanations get two things wrong: they call the camera a tool, and they let agents play both actor and crew. The camera is your observability layer; losing it means losing the single most important governance feature a harness gives you.
| Improv theatre | AI harness | Governance question it answers | Assessed against |
|---|---|---|---|
| Stage manager, rigging, cue systems, safety protocols | The harness: context management, tool routing, permissions, guardrails | Who controls the show while it runs? | ISO/IEC 42001 A.9.2 ISO/IEC 42001 A.6.2.6 NIST AI RMF MANAGE 2.1 |
| Director | The human (Delegation and Description) | Who decides what gets made, and approves what leaves the building? | ISO/IEC 42001 A.3.2 NIST AI RMF GOVERN 3.2 |
| Performers | AI agents; they improvise, they do not follow a fixed script | Where does the non-determinism live? | ISO/IEC 42001 A.6.2.4 NIST AI RMF GOVERN 4.3 |
| Prompt book | System prompt, task specification, policy JSON | What were they told, exactly? | ISO/IEC 42001 A.6.2.2 ISO/IEC 42001 A.6.2.7 NIST AI RMF MEASURE 2.8 |
| Props and set pieces | Tools and resources: APIs, databases, file systems | What can they touch? | ISO/IEC 42001 A.4.4 ISO/IEC 42001 A.4.5 NIST AI RMF MAP 4.1 |
| Cameras and dailies | Observability: logging, telemetry, evaluation (Discernment) | Can every take be reviewed? | ISO/IEC 42001 A.6.2.8 NIST AI RMF MANAGE 4.1 |
| Continuity supervisor | Memory and state management | Does scene 40 remember scene 4? | ISO/IEC 42001 A.7.5 NIST AI RMF MANAGE 4.1 |
| Safety officer at the ready | Human-in-the-loop approval gates | What cannot happen without a human signature? | ISO/IEC 42001 A.9.4 NIST AI RMF MAP 3.5 |
| Call sheet | Orchestration and task queue | Who performs what, in what order? | ISO/IEC 42001 A.6.2.5 NIST AI RMF GOVERN 1.6 |
| "Cut!" | Kill switch: budget limits, loop detection, halt authority | Can you stop a bad take mid-scene? A film studio cannot. A well-built harness can. | ISO/IEC 42001 A.6.2.6 NIST AI RMF MANAGE 2.4 |
About the fourth column. Control identifiers are from ISO/IEC 42001:2023 Annex A and NIST AI 100-1, the AI Risk Management Framework 1.0. The mapping itself is ours, not a certification body’s crosswalk. It exists so a control can be argued for in a budget submission, and it is a starting point for a gap assessment, not a compliance attestation.
Government of Canada readers. Where an agent participates in an administrative decision, these ten controls also map to the Treasury Board Directive on Automated Decision-Making and the Algorithmic Impact Assessment that Directive requires. That mapping is department-specific and depends on your impact level, so it is produced in the readiness assessment rather than asserted here.
Most organizations deploying AI agents today are running Scene 1 with the harness off. The failure is rarely the model; it is the absence of the operating layer around it. Before your agents touch a payments API, a client inbox or a production database, the harness architecture decision has already been made, deliberately or by default.
What happens next: a 45-minute Agentic Workforce Governance Readiness conversation. You leave with a one-page harness gap assessment against the ten controls in the table above. Booking to first output: five business days.
pvanabbema@altnexus.com | altnexus.com | linkedin.com/in/pvanabbema