Software engineering is becoming one of the first fields where agents can carry substantial work for hours. Model capability explains only part of the progress.
Code is unusually agent-ready. It can be executed in a sandbox, compared with tests, reviewed as a diff, released to a small percentage of users, monitored in production, and rolled back to a known working version. Consequential mistakes can be made visible before they become permanent.
Most business workflows lack this machinery. A refund, pricing exception, customer promise, payment instruction, or employee decision may be difficult to test objectively and impossible to reverse after release.
THE MISSING INFRASTRUCTUREEnterprise agents need CI/CD for work—not only for code.
Operational CI makes work testable
Operational continuous integration is the assurance layer around an AI-enabled workflow. Every proposed change to the model, context, tools, instructions, permissions, or policy is tested against the same real work before it receives greater production responsibility.
Scroll to compare →
| Control | Software CI/CD | Operational CI |
|---|---|---|
| Specification | Issue, requirements, and code review standard | Workflow objective, policy, authority, and acceptance criteria |
| Test environment | Local development or sandbox | Synthetic case, shadow mode, or transaction sandbox |
| Automated tests | Unit, integration, and regression tests | Real-task evals, policy checks, and outcome verification |
| Version control | Commit and pull-request history | Versioned prompts, context, skills, tools, policies, and evals |
| Approval | Required reviewers | Human gates for consequential decisions and exceptions |
| Staged release | Canary or percentage rollout | Limited users, cases, authority, value, or transaction volume |
| Observability | Logs, traces, errors, and performance | Full run evidence, interventions, cost, risk, and business outcome |
| Recovery | Rollback to the last working build | Pause, revoke access, reverse actions, and restore prior workflow state |
Define outcome, authority, and unacceptable failure
Run historical, difficult, and refusal cases
Apply identity, policy, and action checks
Keep people at consequential decisions
Release limited responsibility first
Measure outcome and reverse failure
Production evidence becomes the next regression test ↻
Verification determines the autonomy ceiling
The question is not simply whether an agent can perform the task. It is whether the organization can prove that the task was performed correctly and contain the consequences when it was not.
A workflow with objective checks, bounded authority, visible state, and reversible actions can safely increase autonomy faster. A workflow whose result depends on subjective judgment, hidden downstream effects, or irreversible action needs tighter human control.
Observable + Verifiable + Bounded + Reversible = A Higher Safe Autonomy Ceiling
Operational CI in three workflows
Invoice exception
Test source matching, duplicate detection, policy, amount thresholds, approvals, and the prohibition on releasing payment.
Refund request
Replay normal and difficult tickets, verify policy citations, require approval above a threshold, and begin in draft-only mode.
Pricing exception
Check customer tier, margin floor, precedent, approval authority, CRM evidence, and whether the final offer matched the approved terms.
Every production failure should strengthen the test system
A long-tail failure should not end with a corrected transaction. Preserve the trace, understand the missing condition, add the case to the evaluation set, improve the control, and verify that the failure cannot recur after the next release.
This changes evaluation from a launch ceremony into daily operating infrastructure. The organization gains a living body of evidence about which capabilities are safe, under which conditions, and at what level of authority.
THE CIO OPPORTUNITY
Build the assurance system once. Reuse it across workflows.
Common identity, logging, policy enforcement, evaluation, staged-release, and recovery patterns can serve finance, support, revenue, engineering, and other agentic workflows without forcing them into one generic agent.
Safe Scale = Repeatable Verification × Bounded Authority × Fast Recovery
The enterprises that scale agents will not be those willing to take the most risk. They will be those that make agent work easiest to test, observe, constrain, and reverse.