Skip to content

§5.4 Other Agentic Software Factories

DocAble provides a longitudinal account of one engineer developing and operating an agentic software factory. Other organizations begin with different resources and constraints: established engineering teams, existing production processes, accumulated institutional knowledge, and obligations distributed across many people. Their factories may therefore organize delegation, evaluation, and admission differently. The four questions from §5.1 guide the comparison: what the factory produces, what people do, how production is organized, and what tolerances the resulting software must satisfy. The public accounts illuminate different parts of those questions; they need not supply the same answer or document every part equally well.

These are my interpretive reconstructions of public reports, not claims that the companies adopted MAGE. Engineering blogs, talks, and product documentation describe mechanisms and operating experience, but also serve promotional purposes. I use MAGE to interpret the arrangements they describe; the reports do not independently validate the method. Reported measurements retain their scope, and missing information remains missing rather than proving that a mechanism does not exist.

5.4.1 Cloudflare

Cloudflare shows how an organization's policy can become a control over work performed across many teams. Its standards corpus had grown beyond what individual engineers could readily remember. An extraction agent turns each SHOULD or MUST into a structured statement with a stable identifier, so the obligation can be tracked when its containing document changes 1.1. Timo Reimann, “How Cloudflare Enforces Engineering Standards Using AI,” Cloudflare, August 4, 2026, https://blog.cloudflare.com/engineering-standards-enforcement/. That identity supports compliance records and explicit waivers. Standards then meet work at several stages: specification review before implementation, linters during editing, semantic review in CI, and incident review after production. The representation connects a human-authored obligation to machinery that can find and apply it; merely storing the original policy documents would leave more of that connection to each reader.

An approved statement does not automatically block work. Any employee may propose a standard, and a named domain owner approves it. A separate promotion moves it from advisory to enforced; the cited account does not identify who performs that step. For enforced standards, a coordinator agent orchestrates up to seven specialist reviewers, evaluates severity, and can approve a merge request, revoke an earlier approval, or block the merge. Only enforced MUST statements can block, and the rubric favors approval so that a lone warning does not 2.2. Ryan Skidmore, “Orchestrating AI Code Review at Scale,” Cloudflare, April 20, 2026, https://blog.cloudflare.com/ai-code-review/. This separates the decision to adopt an obligation from the decision to let an automated reviewer enforce it.

Human reviewers retain an exception path: commenting “break glass” forces approval and records the override. Cloudflare reports 288 such overrides, 0.6 percent of merge requests, from March 10 to April 9, 2026 2. The override count measures how the control was used. Review runs, findings, and blocked merges similarly describe its operation. The account makes the allocation of authority visible, while the effectiveness of the checking—including violations missed and defects prevented—remains incompletely measured.

Operating experience also changes what later work inherits. The November and December 2025 outages led to rules against calling .unwrap() outside tests and against assuming upstream dependencies were in an expected state. Dangerous configuration patterns can be enrolled in Snapstone's shared deployment machinery. The policy representation changed from Markdown to JSON when agents needed more reliable filtering 3.3. Jeremy Hartman, “Code Orange: Fail Small Is Complete. The Result Is a Stronger Cloudflare Network,” Cloudflare, May 1, 2026, https://blog.cloudflare.com/code-orange-fail-small-complete/. Semantic-review latency also encouraged moving suitable checks into deterministic linters earlier in the workflow 1. These are concrete instances of governance conversion: a failure or recurring cost leads to a reusable change in the environment. The sources describe those changes but do not measure their effect on later incidents.

5.4.2 Spotify

Spotify introduced agents into an established process for changing software across many repositories. Fleet Management selected repositories, ran transformations, opened and merged pull requests, and monitored the resulting deployments, alerting the change author when production failed 4.4. Matt Brown, “Fleet Management at Spotify (Part 3): Fleet-Wide Refactoring,” Spotify Engineering, May 15, 2023, https://engineering.atspotify.com/2023/05/fleet-management-at-spotify-part-3-fleet-wide-refactoring. In 2025, an agent replaced the transformation script at one station while the surrounding targeting, review, and merge infrastructure initially remained the same. The agent addressed changes that deterministic scripts had struggled to cover, including the remaining 30 percent of one fleet migration 5.5. Max Charas and Marc Bruggmann, “1,500+ PRs Later: Spotify’s Journey with Our Background Coding Agent,” Spotify Engineering, November 6, 2025, https://engineering.atspotify.com/2025/11/spotifys-background-coding-agent-part-1. This lets us examine which production structures could be reused and which later required revision.

The late-2025 configuration bounded both the agent's actions and its route to a pull request. It exposed selected Git commands that excluded pushing, shell commands from an allowlist, and a verification tool. Users supplied relevant context in prompts; code-search and documentation tools were withheld 5. The verification tool selected checks from repository contents, and failure prevented a pull request from being opened. The agent could choose how to implement a change within those bounds, but its own assessment could not waive the required checks. After unwanted changes such as disabled tests, Spotify added an LLM judge that compared the diff with the original request. In December 2025, the company reported vetoes in about a quarter of thousands of sessions, with agents correcting their work after about half of those vetoes 6.6. Max Charas and Marc Bruggmann, “Background Coding Agents: Predictable Results Through Strong Feedback Loops (Honk, Part 3),” Spotify Engineering, December 9, 2025, https://engineering.atspotify.com/2025/12/feedback-loops-background-coding-agents-part-3.

A March 2026 report of Spotify's QCon talk describes a further separation between the agent runtime and verification runtime, with pull requests created only after CI checks. It also reports that the judge sometimes rejected valid changes and was later removed as foundation models improved and verification instructions in prompts proved sufficient for its role 7.7. Daniel Curtis, “QCon London 2026: Rewriting All of Spotify’s Code Base, All the Time,” InfoQ, March 18, 2026, https://www.infoq.com/news/2026/03/spotify-honk-rewrite. This is a secondary account of the talk, and it does not imply that the required checks or review disappeared. The sequence shows why maintaining a control includes reassessing both the failures it catches and the valid work it rejects. An engineering response to one generation of agent behavior need not remain economical unchanged.

The inherited process also supports different admission decisions for different producers. Spotify reports more than 2.5 million automated maintenance pull requests, most automatically merged; that cumulative figure chiefly describes the deterministic fleet machinery, not agent-authored work 8.8. Niklas Gustavsson, “Coding Is No Longer the Constraint: Scaling Developer Experience to Teams and Agents at Spotify,” Spotify Engineering, June 3, 2026, https://engineering.atspotify.com/2026/6/code-with-claude-coding-is-no-longer-the-constraint. The QCon account says agent-authored changes still required human review 7. People define migration scope and maintain prompts as tested, version-controlled artifacts. They can also reserve particular decisions: in one migration, the agent was instructed to leave fields unchanged and annotate them whenever a judgment call was needed 59.9. Devon Edwards Joseph, “Background Coding Agents: Supercharging Downstream Consumer Dataset Migrations,” Spotify Engineering, April 22, 2026, https://engineering.atspotify.com/2026/4/background-coding-agents-dataset-migrations-honk-part-4. Delegation therefore varies within the production process rather than being one decision to automate the whole line.

Spotify's later measurements make the limits of that process visible. Its refined rework measure did not rise during the AI ramp-up, while code complexity and pull-request size did; the company kept its thresholds unchanged. Incident retrospectives found no material direct contribution from AI-authored code, yet change volume had grown faster than some verification controls could adapt. One automated dependency upgrade passed its checks and still failed in production 10.10. Tyson Singer, “AI Changed How Spotify Builds. What We Learned (and Fixed) About Quality at Higher Velocity,” Spotify Engineering, September 16, 2026, https://engineering.atspotify.com/2026/9/ai-changed-how-spotify-builds-what-we-learned-and-fixed-about-quality-at-higher-velocity. These observations concern different properties and should not be collapsed into a single quality score. Producing more changes has exposed a capacity problem as well as a check-coverage problem.

5.4.3 Shopify

Shopify's case concerns how work performed in one session becomes knowledge available to later sessions. River, its internal agent, operates in Slack and refuses direct messages. An employee starts work by mentioning it in a channel, leaving a public transcript that colleagues can inspect 11.11. Javier Moreno and Burke Libbey, “Under the River,” Shopify Engineering, May 28, 2026, https://shopify.engineering/under-the-river. This makes the work and its difficulties observable beyond the person who initiated it. It does not ensure that every failure attracts attention, but it provides a shared record from which people can identify missing knowledge instead of depending entirely on a later report.

People use that record to change what later agents receive. They watch River work, identify what it should have known, and add a skill, revise an AGENTS.md file, or record another reusable instruction in the monorepo 11. These artifacts sit alongside code, conventions, intent documents, and runbooks. River's reported merge rate rose from 36 to 77 percent over two months while its foundation model stayed the same 12.12. Tobi Lütke, “Learning on the Shop Floor,” May 9, 2026, https://x.com/tobi/article/2053121182044451016. The improvement occurred with the foundation model unchanged, as people deliberately preserved lessons for later work. The report does not isolate the contribution of those changes. Unlike DocAble's one-engineer account, this mechanism distributes observation and knowledge maintenance across an organization.

The shared knowledge also requires maintenance. Instruction files apply through the monorepo's directory hierarchy, while different agent tools read different files. Shopify describes divergence between those conventions: a directory can contain instructions for one tool and omit the corresponding file for another, so engineers working on the same code supply different context to their agents 1314.13. Tobi Lütke, “Split Brain Problems When Different Team Members Use Different Tools,” X, August 25, 2026, https://x.com/tobi/status/2092259436538495186.14. Amanda Caswell, “Shopify’s CEO Threatened to Ban Claude Code. Anthropic Had Already Closed the Feature Request,” The New Stack, August 25, 2026, https://thenewstack.io/shopify-claude-code-agentsmd/. Automation repairs that drift. Recording a lesson has therefore created another engineering artifact whose continued value depends on keeping it available and consistent. The cost follows from the mechanism that makes the knowledge reusable.

A security-workflow account shows why recording an instruction may be insufficient. River violated a prompt-stated enumeration requirement and acknowledged the omission in its report. Shopify's response was to require rules governing vulnerability closure to be enforced in code. The same account describes draft-only agent output, merge authority held by the engineers responsible for the affected code, a merge queue requiring fresh evidence, and a post-merge check that the default branch contains the fix. Use of that queue rose from about 10 to 80 percent 15.15. Erin Son and Kaiyi Li, “How River Takes Security Work from a Fix to Merge,” Shopify Engineering, September 2, 2026, https://shopify.engineering/river-vulnerability-remediation. That measures adoption of the queue, not the rate at which vulnerabilities escaped. The controls are described for security work; the account does not establish what evidence ordinary River pull requests must carry or how often they fail in production.

5.4.4 Uber

Uber treats the factory itself as something to measure and improve. Its managed agents perform work such as CI repair, review, alert triage, and maintenance; operational signals and schedules can initiate tasks as well as people 16.16. Uday Kiran Medisetty and Adam Huda, “Agentic SDLC at Uber — Building Blocks for Uber’s Software Factory,” AI Engineer World’s Fair, 2026, https://ai.engineer/talks/17-YSUHo6Lk-agentic-sdlc-at-uber-building-blocks-ubers. The relevant question is what it costs to complete useful work through that arrangement. A token reduction, a cheaper foundation model, and a shorter review queue affect different parts of the answer.

Uber adjusts context delivery, tool access, and coordination. Moving deterministic orchestration from repeated agent turns into executable code reportedly cut token use by more than half for common workflows and by more than 90 percent for bulk cases. Real workloads become benchmarks, foundation models are selected against cost and reliability, and managed agents are evaluated in units such as cost per merged pull request, review, alert, or cleanup 1716.17. Uday Kiran Medisetty, “Running a Software Factory Efficiently at Uber Scale,” Uber Engineering, 2026, https://www.uber.com/us/en/blog/efficient-software-factory/. Those units connect resource use to an outcome rather than counting agent activity alone. They also make the surrounding engineering choices visible: some work is cheaper when executed directly than when repeatedly reconstructed through model interactions.

Evaluation begins inside the producing agent's loop, with static-analysis repair, simulator screenshots compared with design specifications, and staging integration. The resulting pull request carries a table of checks and evidence assembled by that agent. The account does not say that independent machinery re-derives every entry 16. Additional model-based review, CI, and human approval follow. Uber describes a possible move toward partial automatic approval, but does not publish the qualifying evidence or eligibility criteria 18.18. Will Bond and Ameya Ketkar, “Building uReview, Uber’s Multi-Agent Code Review Engine,” AI Engineer World’s Fair, 2026, https://ai.engineer/talks/EL123UNokkI-building-ureview-ubers-multi-agent-code-review. The reported review layers therefore tell us how evaluation is arranged without establishing the assurance of every proposed change.

Reviewer attention has its own capacity limit. Time to first review rose from three hours in 2024 to nine in 2026 as volume and change size grew 18. Uber schedules maintenance jobs on Sunday to use spare CI capacity, but separately caps the diffs engineers will face on Monday 16. This is a concrete process-design decision: spare compute can produce more work than the next human stage can absorb. The organization also tracks revert rate, F1, and time to recovery alongside production cost, although the cited accounts do not publish a complete outcome series.

5.4.5 Amazon

Kiro supplies a reusable workflow for making intent explicit before implementation. A feature specification records user stories and acceptance criteria, develops an architectural design, and divides implementation into tasks that engineers can review 19.19. Kiro, “Specs,” Kiro Documentation, 2026, https://kiro.dev/docs/specs/. In MAGE's vocabulary, these requirements and design artifacts are engineering models. DocAble arranged related work through project-specific work orders and design documents; Kiro makes a specification workflow part of the harness. Adopting it still leaves a team responsible for the content of those artifacts and their connection to its system.

The workflow can connect specifications to evidence and enforcement. Kiro's IDE can extract properties from requirements and generate property-based tests, whose execution is optional by default and whose results are explicitly distinguished from proof 20.20. Kiro, “Correctness with Property-based tests,” Kiro Documentation, 2026, https://kiro.dev/docs/specs/correctness/. Configured hooks can run checks after edits or block tool calls when preconditions fail 21.21. Kiro, “Hooks,” Kiro Documentation, 2026, https://kiro.dev/docs/hooks/. Engineers must decide which properties matter, whether the tests express them correctly, and where failure should prevent work from proceeding. A supplied workflow makes these connections possible; its presence does not establish that a particular project enabled them.

Amazon's Bedrock account makes some of that adopting-team work visible. Six senior engineers rebuilt part of the inference engine in 76 days, using small modules, reviewable commit-sized tasks, focused operational tools, tests, end-to-end canaries, and senior inspection of generated code 22.22. Kiro, “Six engineers completed a 30-person project in 76 days,” Amazon Web Services, accessed November 6, 2026, https://kiro.dev/case-studies/bedrock/. AWS reports a twentyfold increase in normalized commit velocity, from two to forty commits per developer per week 23.23. Swami Sivasubramanian, “How frontier teams are reinventing AI-native development,” Amazon Web Services, June 10, 2026, https://aws.amazon.com/blogs/machine-learning/how-frontier-teams-are-reinventing-ai-native-development/. That is a production-activity measure. It does not establish defect reduction or separate Kiro's contribution from team selection and workflow redesign.

Amazon also reports experiments on improving Kiro itself. Kiro-trace proposes prompt or tool changes from recorded runs. A separate deterministic gate screens targeted failures, checks regression assertions, and compares candidates with an unchanged control on the full benchmark. It requires gains on selected tasks without losses beyond a specified margin elsewhere; a person decides whether to merge. One apparent improvement was rejected after it proved to be a lucky run. Evaluation mistakes became checks, and confirmed fixes add regression assertions 24.24. Myeongsoo Kim et al., “How Kiro fixes itself, and how we keep the fixes honest,” Kiro, October 20, 2026, https://kiro.dev/blog/how-kiro-fixes-itself/. These internal experiments illustrate the same engineering problem at another level: improving the fabricator requires evidence about the improvement and maintenance of the machinery that evaluates it.

5.4.6 GitLab

GitLab illustrates how an integrated development workflow can bring existing organizational knowledge and controls to agent work. Repository instructions, chat rules, review skills, CI configuration, and workflow definitions supply different kinds of context 25.25. GitLab, “Customize GitLab Duo Agent Platform,” GitLab Documentation, 2026, https://docs.gitlab.com/user/duo_agent_platform/customize/. The documented security-review flow can report findings but cannot approve a change 26.26. GitLab, “Security Review Flow,” GitLab Documentation, 2026, https://docs.gitlab.com/user/duo_agent_platform/flows/foundational_flows/security_review/. For human-triggered flows running on runners, composite identity limits access to the intersection of the service account’s and triggering user’s privileges 27.27. GitLab, “Composite Identity,” GitLab Documentation, n.d., https://docs.gitlab.com/user/duo_agent_platform/composite_identity/. Tool-governance rules can also constrain actions; project rules may tighten but not loosen a group rule 28.28. GitLab, “Agent Tool Governance,” GitLab Documentation, n.d., https://docs.gitlab.com/user/ai-governance/tool-governance/. These are distinct functions: instructions inform the work, permissions restrict actions, and the approval boundary reserves a consequential decision.

GitLab's internal playbook also makes acceptance depend on the surrounding pipeline. Its failing-test-first practice is backed by a CI job that fails when the test count decreases unless an explicit bypass label is present, addressing agents that remove tests to obtain a passing result 29.29. GitLab, “AI-Assisted Development Playbook,” GitLab Handbook, 2026, https://handbook.gitlab.com/handbook/engineering/workflow/ai-assisted-development/. A test-count check cannot establish that the remaining tests express the right requirements. It does make one recurring way of evading them consequential at admission.

5.4.7 What the Factories Share, and Where They Differ

The four factory questions reveal different arrangements rather than a common maturity sequence. The organizations differ in where they place engineering judgment, what they permit agents to decide, and which decisions remain with people or independent controls. Their measurements answer different questions: override frequency, veto and correction rates, merge rate, token cost, review delay, and engineering-health indicators cannot be combined into a single ranking of factory effectiveness.

One important difference is where the authority to accept work resides. Cloudflare distinguishes standards that advise reviewers from those whose violations can block a merge, while retaining a recorded human override 2. Spotify inserted agents into an existing production process but kept required verification and, in the reported configuration, human review of agent-authored changes 57. Shopify’s security workflow reserves merge authority to the engineers responsible for the affected code 15; GitLab’s documented security-review flow can report findings but cannot approve a change 26. These arrangements reserve admission authority in different ways, but none treats the producing agent’s claim of completion as sufficient authority to admit consequential changes. Process design determines who or what may make the admission decision; tolerance and its evidence determine what that decision must establish.

Two recurring costs cut across those arrangements. The first is supervisory capacity. Uber's Sunday maintenance schedule would be incomplete without its separate limit on Monday's review load. Spotify's ability to generate more changes exposed limits in verification even when its measured rework did not rise. A factory therefore has to organize evaluation and admission alongside realization. Increasing the rate of one stage can shift delay or risk to another; the production system must be assessed through the work it can accept and sustain.

The second cost is maintaining what later work inherits. Shopify repairs divergent instructions; Cloudflare revises policy representations and moves suitable checks earlier; Kiro's improvement experiments acquire new regression assertions. Each preserved lesson can reduce future reconstruction or repeated judgment, but the resulting artifact must remain connected to the work it governs. This is the carrying-cost question from §5.1 in operational form. Improvements in foundation models can also change that calculation, as Spotify's retired judge illustrates; Shopify's rising merge rate with an unchanged model shows that the surrounding factory can improve independently.

Works Cited

  1. Reimann, Timo. “How Cloudflare Enforces Engineering Standards Using AI.” Cloudflare, August 4, 2026. https://blog.cloudflare.com/engineering-standards-enforcement/.
  2. Skidmore, Ryan. “Orchestrating AI Code Review at Scale.” Cloudflare, April 20, 2026. https://blog.cloudflare.com/ai-code-review/.
  3. Hartman, Jeremy. “Code Orange: Fail Small Is Complete. The Result Is a Stronger Cloudflare Network.” Cloudflare, May 1, 2026. https://blog.cloudflare.com/code-orange-fail-small-complete/.
  4. Brown, Matt. “Fleet Management at Spotify (Part 3): Fleet-Wide Refactoring.” Spotify Engineering, May 15, 2023. https://engineering.atspotify.com/2023/05/fleet-management-at-spotify-part-3-fleet-wide-refactoring.
  5. Charas, Max, and Marc Bruggmann. “1,500+ PRs Later: Spotify’s Journey with Our Background Coding Agent.” Spotify Engineering, November 6, 2025. https://engineering.atspotify.com/2025/11/spotifys-background-coding-agent-part-1.
  6. Charas, Max, and Marc Bruggmann. “Background Coding Agents: Predictable Results Through Strong Feedback Loops (Honk, Part 3).” Spotify Engineering, December 9, 2025. https://engineering.atspotify.com/2025/12/feedback-loops-background-coding-agents-part-3.
  7. Curtis, Daniel. “QCon London 2026: Rewriting All of Spotify’s Code Base, All the Time.” InfoQ, March 18, 2026. https://www.infoq.com/news/2026/03/spotify-honk-rewrite.
  8. Gustavsson, Niklas. “Coding Is No Longer the Constraint: Scaling Developer Experience to Teams and Agents at Spotify.” Spotify Engineering, June 3, 2026. https://engineering.atspotify.com/2026/6/code-with-claude-coding-is-no-longer-the-constraint.
  9. Edwards Joseph, Devon. “Background Coding Agents: Supercharging Downstream Consumer Dataset Migrations.” Spotify Engineering, April 22, 2026. https://engineering.atspotify.com/2026/4/background-coding-agents-dataset-migrations-honk-part-4.
  10. Singer, Tyson. “AI Changed How Spotify Builds. What We Learned (and Fixed) About Quality at Higher Velocity.” Spotify Engineering, September 16, 2026. https://engineering.atspotify.com/2026/9/ai-changed-how-spotify-builds-what-we-learned-and-fixed-about-quality-at-higher-velocity.
  11. Moreno, Javier, and Burke Libbey. “Under the River.” Shopify Engineering, May 28, 2026. https://shopify.engineering/under-the-river.
  12. Lütke, Tobi. “Learning on the Shop Floor.” May 9, 2026. https://x.com/tobi/article/2053121182044451016.
  13. Lütke, Tobi. “Split Brain Problems When Different Team Members Use Different Tools.” X, August 25, 2026. https://x.com/tobi/status/2092259436538495186.
  14. Caswell, Amanda. “Shopify’s CEO Threatened to Ban Claude Code. Anthropic Had Already Closed the Feature Request.” The New Stack, August 25, 2026. https://thenewstack.io/shopify-claude-code-agentsmd/.
  15. Son, Erin, and Kaiyi Li. “How River Takes Security Work from a Fix to Merge.” Shopify Engineering, September 2, 2026. https://shopify.engineering/river-vulnerability-remediation.
  16. Medisetty, Uday Kiran, and Adam Huda. “Agentic SDLC at Uber — Building Blocks for Uber’s Software Factory.” AI Engineer World’s Fair, 2026. https://ai.engineer/talks/17-YSUHo6Lk-agentic-sdlc-at-uber-building-blocks-ubers.
  17. Medisetty, Uday Kiran. “Running a Software Factory Efficiently at Uber Scale.” Uber Engineering, 2026. https://www.uber.com/us/en/blog/efficient-software-factory/.
  18. Bond, Will, and Ameya Ketkar. “Building uReview, Uber’s Multi-Agent Code Review Engine.” AI Engineer World’s Fair, 2026. https://ai.engineer/talks/EL123UNokkI-building-ureview-ubers-multi-agent-code-review.
  19. Kiro. “Specs.” Kiro Documentation, 2026. https://kiro.dev/docs/specs/.
  20. Kiro. “Correctness with Property-based tests.” Kiro Documentation, 2026. https://kiro.dev/docs/specs/correctness/.
  21. Kiro. “Hooks.” Kiro Documentation, 2026. https://kiro.dev/docs/hooks/.
  22. Kiro. “Six engineers completed a 30-person project in 76 days.” Amazon Web Services. Accessed November 6, 2026. https://kiro.dev/case-studies/bedrock/.
  23. Sivasubramanian, Swami. “How frontier teams are reinventing AI-native development.” Amazon Web Services, June 10, 2026. https://aws.amazon.com/blogs/machine-learning/how-frontier-teams-are-reinventing-ai-native-development/.
  24. Kim, Myeongsoo, Patrick Chapman, Chang Liu, et al. “How Kiro fixes itself, and how we keep the fixes honest.” Kiro, October 20, 2026. https://kiro.dev/blog/how-kiro-fixes-itself/.
  25. GitLab. “Customize GitLab Duo Agent Platform.” GitLab Documentation, 2026. https://docs.gitlab.com/user/duo_agent_platform/customize/.
  26. GitLab. “Security Review Flow.” GitLab Documentation, 2026. https://docs.gitlab.com/user/duo_agent_platform/flows/foundational_flows/security_review/.
  27. GitLab. “Composite Identity.” GitLab Documentation, n.d. https://docs.gitlab.com/user/duo_agent_platform/composite_identity/.
  28. GitLab. “Agent Tool Governance.” GitLab Documentation, n.d. https://docs.gitlab.com/user/ai-governance/tool-governance/.
  29. GitLab. “AI-Assisted Development Playbook.” GitLab Handbook, 2026. https://handbook.gitlab.com/handbook/engineering/workflow/ai-assisted-development/.
© James C. Davis, 2026–present