AI Agent Engineering Task Flow: Cited Swimlane (2026)
Coding agents have evolved from autocomplete to whole-task execution: take an issue, produce a pull request, iterate on test failures, and submit for human review. The agent loop is non-deterministic; the process around it (issue intake, CI, code review, merge) is BPMN-deterministic. The canonical pattern wraps the agent loop as a bpmn:callActivity and keeps the wrapper process correct.
Two named case studies
Cognition's Devin (released in technical preview March 2024, generally available 2024 to 2025) is the most-cited engineering coding agent, and Cognition has published its own internal use of it. The account is unusually concrete about where the gate sits: “Tag @ Devin with a request and we get a PR we can review and test out”, and every pull request Devin posts carries a review link so the diff is one click away. Anthropic's Claude Code (released as a research preview February 2025, generally available May 2025) is the second canonical example; its documentation describes the same wrapper process, a tool that “reads your codebase, edits files, runs commands” and then “stages changes, writes commit messages, creates branches, and opens pull requests”.
Both deployments place the human review at the pull request, not inside the loop. The loop is allowed to iterate freely on test failures; the human reviews only the final candidate. The pattern matches Anthropic's “Building Effective Agents” (Schluntz 2024) workflow-vs-agent distinction: the agent loop is a service step inside a workflow, not a workflow itself.
A third example is often named alongside these two: GitHub Copilot Workspace, the GitHub Next experiment that turned an issue into a planned pull request. It is not a live deployment and should not be cited as one. Its preview endpoint at copilot-workspace.githubnext.com returns a 404, and the Copilot Workspace user manual repository is archived, with its last commit on 2 September 2025 (both checked October 2026). The issue-to-pull-request shape it demonstrated continues in GitHub's current coding-agent products, but those are different products and need their own citation before they appear here.
Where the human gates sit
Two gates. The first is the brief (human framing of the task before the agent starts); the second is the PR review before merge. The PR review is the strict gate; merge into main is the irreversible action and warrants human sign-off regardless of CI green. Both Devin and Claude Code expose this gate explicitly in their documentation.
Where the handoffs sit
Three handoffs. Human-to-agent at the brief (sequence flow into the agent lane). Agent-to-CI at PR creation (sequence flow into the system lane). CI-to-human or back-to-agent at the test result (gateway-routed sequence flow). All within one pool because the engineer, the agent, and the CI are inside the same organisation; an external CI provider would push the CI step into a separate pool with a message flow.
Workforce-impact note
Coding agents have not (as of April 2026) replaced engineering roles at measurable scale; published deployments report task-time savings on well-bounded tickets and increased throughput per engineer. The rep-as-reviewer pattern shifts senior engineering time from writing to reading. For the methodology, see aijobimpactcalculator.com.
Related pages
- BPMN call-activity : the spec section that licenses the wrapper.
- Human vs agent swimlanes
- buildingeffectiveagents.com : the engineering reference for what goes inside the loop.