An agent's behavior changed on a Tuesday. Nobody merged a commit. The repository was untouched, the deploy log was empty, and the output was different anyway.

That is the operating consequence of agentic delivery that buyers usually discover late: the thing you approved is not the thing you can redeploy.

What this edition claims, and what it does not

No benchmark, error rate, incident count or vendor comparison is asserted here. The framework sources cited below are voluntary guidance rather than certifications, and no conformity determination is drawn from any of them.

What follows is this publication's analysis of a change-control problem, plus framework material and one vendor description, both listed at the end.

The release unit is not the commit

This section is analysis.

For deterministic software, the release unit is the build artifact. Version it, deploy it, and roll it back by redeploying the previous one. The model of the world is simple enough that change control has a single object.

An agent has at least five inputs to behavior, and they version independently. The model version. The system instructions and configuration. The tool schemas and what those tools return. The permission scope the agent holds. And the evaluation set used to judge whether it behaved acceptably.

Any one of those can move without a code commit. A hosted model gets updated upstream. An instruction gets edited in a console by someone who never opens the repository. A tool adds a field and the agent starts treating it as a signal. A credential scope gets widened during an incident and is never narrowed back.

The commit is a real artifact. It is just not the one that determines what the agent does.

Why rollback does not restore behavior

This section is analysis. Rollback works because the artifact is versioned and the surrounding environment is assumed constant. Agentic delivery breaks that assumption in four places.

Pinned is not permanent. A pinned model version is a dependency with its own lifecycle. When a hosted version is retired, the rollback target no longer exists, and what the fallback produces is a new release rather than a restoration. Teams discover this at the moment they most need the old one.

Configuration lives outside version control more often than anyone admits. Instructions, sampling settings and tool enablement are frequently editable in a console. A console edit is not reviewed by whoever reviews pull requests, and it is not in the audit trail reviewers believe they have.

The evaluation set ages. A regression suite for generative behavior is a corpus of cases with expected properties, and its relevance decays as real requests drift. It can pass cleanly while the population it was built for has moved somewhere the suite never covered.

Permission drift persists quietly. A scope widened to get through an incident is a permanent behavior change that no commit records and no test catches.

What the framework material actually gives you

CISA's Secure by Design guidance frames ownership plainly, stating that every technology provider must take ownership at the executive level to ensure their products are secure by design, and that the burden has been placed disproportionately on the shoulders of consumers and small organizations and away from the producers of the technology. It also publishes a Secure by Demand guide described as concise and usable by any customer of software during procurement discussions with third party resellers or service providers.

That guide is written about buying software, and applying its logic to an agent your own team assembled is this publication's reading rather than CISA's claim. The reading is narrow: the questions you would put to a vendor about change handling are the questions to put to your own build team.

NIST Special Publication 800-218, the Secure Software Development Framework, describes a core set of high-level secure software development practices that can be integrated into each SDLC implementation. Read narrowly, that gives an organization vocabulary for reviewing how a system changes over time. It does not tell you what an agent release is, and no practice identifier or conformity claim is made here.

OWASP's Top 10 for Large Language Model Applications names risk categories that land directly on this problem, including prompt injection and excessive agency. Excessive agency is a statement about scope, and scope is one of the five inputs. That makes it a release-control question before it is a security-control question.

Sandboxing is being productized, which tells you something

OpenAI published a description on November 18, 2026 of how Codex runs commands on a machine while enforcing filesystem and network limits the user controls, across macOS, Linux and Windows. That is the extent of what this edition reviewed: the announcement summary. No enforcement detail, bypass test or efficacy result is characterized here.

Read carefully, it is a useful signal rather than a solution. The containment boundary is being treated as a product feature with customer-set limits, which is the right direction. It also draws the line worth understanding: a sandbox constrains where an action can land. It does not establish that the action was correct, and it does not record which set of instructions produced it.

Three placements for a release record

When the bundle has to be identifiable, there are three places to put the record, and they fail differently.

  • In the repository. Instructions, tool schemas and model pins are versioned beside code. Cleanest review story available and the easiest to explain to an auditor. It fails the moment a console stays writable, because the repository becomes a partial truth told confidently.
  • In a policy gate. Every behavior-affecting change must pass a single approval path regardless of origin. Strongest enforcement, and its failure mode is the emergency bypass: an urgent change that skips the gate and is never logged as a release, which is precisely the change you will need to explain later.
  • In the output itself. Each result carries an identifier resolving to the bundle in force at generation time. Best forensic position, and the hardest to retrofit, because it touches every call path and every consuming system.

None is correct in the abstract. The choosing question is what has to be true when someone asks, six months from now, why the agent said that. Only one of the three answers with a record instead of a recollection.

Four fields a release record needs

  • A bundle identifier. Not a commit hash. Something that resolves to all five inputs together.
  • The model version in force, and the stated fallback. Including what happens when the pinned version is retired, and who decided.
  • The instruction and tool-schema versions, with the location of approval. The console path counts here; if a setting can change there, that is where the approval has to be recorded.
  • The evaluation set version and its last run date. A pass with an unknown date is not evidence of anything.

Five questions before the next deploy

  • What can change this agent's behavior without a commit, and who is allowed to do it? If the answer includes people who do not know they are changing a release, start there.
  • If the pinned model were retired next month, what is the plan? And is that plan a tested release, or an intention?
  • Is every console-editable setting mirrored in version control, or is the console authoritative? Whichever it is, say so out loud to the people signing off.
  • When permissions were last widened under pressure, was that recorded as a change? Urgency is a legitimate reason to widen scope. It is not a reason to forget it.
  • Does an output carry enough to reconstruct which bundle produced it? If reconstruction requires asking a person, the answer depends on that person still being there.

A 30-day check

  • Days 1 to 5. List every input that can change agent behavior. For each, record the owner and the system where it actually lives, not where it is supposed to live.
  • Days 6 to 10. Take one production output and reconstruct the bundle behind it using documents only. Note every point where you had to ask someone.
  • Days 11 to 15. Count console or administrative edits that changed behavior without a commit over two weeks. If the count cannot be produced, that is the finding.
  • Days 16 to 22. Re-run the evaluation set and compare against its last run. Record the date gap as plainly as the result.
  • Days 23 to 30. Write the model-retirement plan for your most-used pinned version, and name who owns executing it.

The productivity claim, stated precisely

This section is analysis.

The case for agentic delivery is real, and it moves work rather than removing it. Fewer hours producing the change, more hours deciding what acceptable behavior means and who signs for it. That trade is usually worth making.

The failure is narrower than a productivity failure. A team with no identifiable release unit gets speed it cannot bank, because every incident turns into archaeology and every audit turns into a reconstruction project. The delivery velocity is genuine; the confidence it should have produced is not.

For a buyer, that reframes the question. It is not whether the agents are fast enough. It is whether the speed survives the first serious question about what changed.

What remains unconfirmed

  • Any enforcement detail, bypass test or efficacy result for the Codex sandbox limits described in OpenAI's November 18, 2026 summary. Only the summary was reviewed.
  • Whether any specific platform versions prompts, tool schemas and configuration as first-class, independently rollbackable artifacts. No vendor documentation was reviewed for this edition.
  • Any measured effect of release-record practices on incident duration, rollback success or audit outcomes. None is asserted.
  • Retirement timelines or notice periods for any hosted model version. None were reviewed, and no specific retirement is claimed.
  • Current benchmarks or accepted practice for generative regression coverage. None were reviewed.
  • Any customer outcome attributed to agentic delivery or vibe coding. None is reported here.
  • Market data. No stock, index or valuation figures were reviewed for this edition.

Evidence notes and limits

CISA's Secure by Design page is the source for the statement that every technology provider must take ownership at the executive level to ensure its products are secure by design; for the statement that the burden has been placed disproportionately on the shoulders of consumers and small organizations and away from the producers of the technology; and for the existence and described use of the Secure by Demand guide in procurement discussions with third party resellers or service providers. It is voluntary guidance rather than a certification, and the article's application of a software-procurement guide to an internally built agent is labeled as this publication's reading.

NIST Special Publication 800-218, the Secure Software Development Framework, is cited only for its description of a core set of high-level secure software development practices that can be integrated into each SDLC implementation. No practice identifier is asserted and no conformity claim is made.

OWASP's Top 10 for Large Language Model Applications is cited only for the existence of its named risk categories, including prompt injection and excessive agency. No ranking, score, version identifier or claim about any product's vulnerability is asserted.

OpenAI's Codex sandbox post, dated November 18, 2026, is cited only for its announcement-summary description: how Codex runs commands on a machine while enforcing filesystem and network limits the user controls across macOS, Linux and Windows. No enforcement mechanism, bypass claim, test result or customer outcome is characterized.

No interface specification, configuration export, access log, incident record, evaluation corpus or customer environment was reviewed. The five-input framing, the four rollback failure modes, the three placements, the four fields, the five questions and the 30-day check are this publication's analysis. Nothing here establishes that any organization lost a rollback target, shipped an unreviewed console change, or changed a procurement decision.

The advertisement above is separate from this analysis and is not a source for any claim in it.