A working demo proves a model can produce a plausible result once. It does not prove the system will behave when an agent holds credentials, calls tools and changes production.
That distinction is not academic. It is the difference between a prototype and a deliverable, and a NIST forum session description now gives it a useful name.
What NIST put on the record
A September 22, 2026 session at NIST's SSCA Forum, presented by John Willis of AGCP.ai, was described this way: it "focused on how the inclusion of large language models with agents in the software supply chain shifts software assurance problems to execution assurance problems."
That is a session description, not a standard. It is still useful, because it names where the risk moves. Software assurance asks whether the code is sound. Execution assurance asks what happened when the agent ran: which tool it called, with what authority, on whose behalf, and with what record.
NIST's NCCoE is working the adjacent problem. Its agentic AI identity and authorization project notes that organizations are deploying agents for tasks "ranging from information retrieval and workflow automation to software development and cybersecurity operations." It then states the failure list plainly: "Without strong identity, authorization, and governance, organizations risk data leaks, compliance failures, prompt injection, and unpredictable autonomous behavior."
Read those two sources together and the buyer's problem comes into focus. One describes the shift from code review to runtime review. The other describes the controls that have to exist before runtime review is possible.
The demo-to-production gap is an evidence gap
This section is analysis.
Unstructured prompt-first delivery optimizes for the next working output. That is rational during discovery. The trouble starts when the output becomes a production change, because the artifact that made the demo work is usually a prompt, a model version and a manually adjusted context window. None of those is the artifact a reviewer can audit six months later.
Three things separate a demo from a deliverable. A stated behavior: what the agent may and may not do. An evaluation set: tasks with known correct outcomes, including tasks the agent should refuse. An execution record: the inputs, tools, parameters, outputs and approvals that let someone reconstruct the run without interviewing the engineer who built it.
NIST's Secure Software Development Framework supplies a vocabulary for the first and third. NIST's draft handbook describes the SSDF as organized into four categories: Prepare the Organization, Protect the Software, Produce Well-Secured Software, and Respond to Vulnerabilities. It also notes that the SSDF does not prescribe implementation but focuses on intended outcomes.
That is the right way to use it. The categories are a checklist for questions, not a certificate. An agent that writes code or operates a tool is now inside the development lifecycle, which means its prompts, permissions and evaluation sets are in scope rather than adjacent to it.
Why the fifth integration is where prompt-first stalls
This section is analysis.
A single agent in a single tool is easy to demo. The fifth integration introduces identity, failure handling and data boundaries. The agent now has a credential. It can reach a system that holds customer records or moves money. It can fail in a way that looks like a success message.
NCCoE's risk list becomes operational at that point. A data leak is not a model-quality problem. A compliance failure is not a prompt-engineering problem. Prompt injection is not solved by asking the model nicely. Unpredictable autonomous behavior is what a security reviewer means when they ask who can stop the agent.
An agent with a broad credential is not a productivity feature. It is a new privileged account, and it needs the same questions any privileged account gets: who owns it, what it can reach, what it records, and how it is revoked.
Three delivery models, and what each one costs
- Prompt-first pilot. Fastest to a demo, lowest initial cost, highest cost of proof. Appropriate for discovery and internal experiments. A poor fit for a change that touches money, customers or regulated data.
- Framework-first delivery. The team writes the behavior spec, the evaluation set, the permission model and the rollback before the agent runs unattended. Slower to first demo, faster to a reviewable release. It maps to the SSDF categories without claiming certification.
- Gated agent. The agent can act, but only inside a scope with a human approval step and a recorded decision. The middle path, and often the one that survives a security review because the evidence already exists.
None of these is universally correct. The choosing question is not which model is smarter. It is which model lets your organization prove what happened.
What a buyer should require before signing
- Named behavior and an enforcement point. What the agent may do, what it may not do, and which system enforces the boundary. "It is in the prompt" is not enforcement.
- An evaluation set with refusal cases. Not just tasks it should complete. Tasks it should decline, and a record of how it declined.
- An execution record. Inputs, tools called, parameters, outputs, approvals, model and version. Reconstructable without a meeting.
- Rollback and ownership. Who can stop the agent, who owns the patch, and what happens when the underlying model changes.
- Reviewer capacity. Who reads the agent's output, what they are accountable for, and how many changes they can actually review in a week.
The review queue is a capacity problem
This section is analysis.
When generation gets cheap, the volume of generated change rises faster than reviewer headcount. The scarce resource stops being the ability to produce a change and becomes the ability to prove one is correct.
That is why a delivery model that ignores review capacity tends to fail late. The demo is impressive. The backlog is not. The security review arrives with three hundred changes and no evaluation set, and the team discovers that the constraint was never the model.
NIST's draft handbook includes a section on continuous monitoring as part of its manager's guide. The observation that an agent in production is a running system rather than a finished release is this publication's analysis.
A 30-day check
- Days 1 to 5. List every agent or AI-assisted workflow currently touching a production system, and the credential each one uses.
- Days 6 to 10. For each, record which tools it can call and who approved that scope.
- Days 11 to 15. Build a 20-task evaluation set with known outcomes, including five tasks the agent should refuse.
- Days 16 to 22. Pick three past agent actions and reconstruct them from logs alone. Note which artifact was missing.
- Days 23 to 30. Write the stop rule: who can halt the agent, what triggers a halt, and who reviews its output each week.
What remains unconfirmed
- Any measured defect rate, delivery speed or cost effect for agent-generated code. No benchmark or customer result was reviewed, and none is asserted.
- Whether any specific vendor's runtime governance implementation meets the NIST framing. Only the framing was reviewed, not a product.
- Whether the NCCoE agentic AI identity project has published implementation guidance. The page reviewed describes the project and a concept paper response count, not a final practice guide.
- Whether SSDF categories are sufficient for agentic systems. They are cited as vocabulary, not as a validated control set for agents.
- Market data. No stock, index or valuation figures were reviewed for this edition.
Evidence notes and limits
NIST's CSRC presentation page for "Beyond Guardrails: Runtime Execution Governance for Trustworthy Agents in the Software Supply Chain," dated September 22, 2026, is the source for the presenter, John Willis of AGCP.ai, the SSCA Forum context, and the quoted statement that the session focused on how the inclusion of large language models with agents in the software supply chain "shifts software assurance problems to execution assurance problems." It is a presentation description and is treated as a design position, not a NIST policy, standard or certification. No outcome is drawn from it.
NIST's NCCoE project page on agentic AI identity and authorization is the source for the description of agent tasks including information retrieval, workflow automation, software development and cybersecurity operations; for the quoted statement that "Without strong identity, authorization, and governance, organizations risk data leaks, compliance failures, prompt injection, and unpredictable autonomous behavior"; for the February 2026 concept paper and more than 600 responses; and for the plan to produce an SP 1800 series practice guide. The page describes a project in progress, and no final guidance is asserted.
NIST's publication page for SP 800-218, Secure Software Development Framework Version 1.1, is the source for its February 2022 publication date and for the description of the SSDF as a core set of high-level secure software development practices that can be integrated into each SDLC implementation and used as a common vocabulary between producers and purchasers. It is guidance rather than a certification, and no conformity determination is drawn from it.
NIST's initial public draft of SP 800-100 Revision 1, Information Security Handbook: A Guide for Managers, is the source for the SSDF's four categories, for the statement that the SSDF does not prescribe implementation but focuses on intended outcomes, for the AI Risk Management Framework Core functions Govern, Map, Measure and Manage, for the discussion of CSF 2.0, and for the presence of a continuous monitoring section. It is a draft, not final guidance, and no claim is made that it speaks specifically to agentic systems.
No vendor implementation, audit, contract, deployment record, evaluation set or third-party test was reviewed. The demo-to-production framing, the three delivery models, the five buyer requirements, the review-queue observations and the 30-day check are this publication's analysis. Nothing here establishes that any organization reduced defects, improved delivery speed or prevented an incident.
The advertisement above is separate from this analysis and is not a source for any claim in it.