One bounded workflow, from source work item to writeback.
Every automation begins from a real work item in your tracker and ends with a recorded writeback. In between, the platform runs explicit phases with attempts, artifacts, and governed handoffs.
Eight concepts that describe everything the platform does.
The business-facing item in Azure DevOps, GitHub Issues, or GitLab Issues. The agent never invents work. Every run is anchored to an item your team already tracks.
An execution-facing record linked to the source. Process state, phases, attempts, and artifacts live here so the source stays clean for product owners.
A trigger event detected on the source: a new comment, a status change, a review request, or manual action. Episodes start or extend a job.
A bounded unit of agent work: refinement, implementation, review-fix. Each phase has explicit inputs, a model provider, and one or more attempts.
A single execution of a phase. Attempts are retryable and recorded. If an attempt fails, you see why and can repair, retry, or escalate.
Inspectable evidence produced by a phase: refined description, branch, commit, PR link, comment, or writeback result.
The platform updates the source work item with status, links, and comments, closing the loop with the system your team already trusts.
Every episode, attempt, artifact, provider call, and human decision is recorded for portal inspection and later export.
Source work item, to Agent Job, to writeback. Visible at every step.
A source work item exists
Every automation starts from a real business-facing item, initially an Azure DevOps User Story. Nothing in the platform happens without one.
scroll to advance · 6 stages remaining
Seven stages, each with explicit inputs, outputs, and audit.
- Stage 01 · Sourceplatform
A real work item enters the platform.
The configured work-item connection is observed through polling or authenticated webhooks. When an item matches routing and enablement policy, an episode is created.
- Routing maps source work to a repository
- Enablement policy is explicit
- First episode opens the Agent Job
- Stage 02 · Orchestrationplatform
An Agent Job is created, linked, and scheduled.
The Agent Job carries process state separate from the source story. Product owners read the source; operators read the job.
- Linked back to source
- Process graph initialized
- Provider routing decided per phase
- Stage 03 · Refinementplatform
The first phase produces a refined description.
Refinement reads the source, asks for clarifications when ambiguous, and writes a structured description back as an artifact.
- Bounded by phase config
- Provider call recorded
- Refined-description artifact attached
- Stage 04 · Implementationplatform
A branch and commits appear in your repository.
Implementation runs on the routed repository using configured model credentials. The agent writes code, validates, and commits to a job-scoped branch.
- Branch: agent/<work-item-id>
- Commits attributed to the job
- Bootstrap checks re-validated
- Stage 05 · Pull requestplatform
A PR opens - your reviewers, your CI, your rules.
The pull request is a normal PR. CI runs, branch protections apply, and code owners are pinged.
- PR links back to source
- CI status surfaced
- Awaiting review is a human checkpoint
- Stage 06 · Review fixhuman
Review comments become the next episode.
When a reviewer leaves a review comment, the platform creates a follow-up episode. A review-fix phase runs against the same Agent Job.
- Threaded back to original PR
- New attempts under same job
- Reviewer can dismiss or escalate
- Stage 07 · Writebackaudited
Status, links, and comments flow back to the source.
When the PR merges or the job closes, writeback updates the source work item and records every artifact and decision.
- Source work item updated
- PR and branch links attached
- Audit record preserved
Three concrete jobs: happy path, review-fix loop, and governed repair.
Every run below is a real shape of audit timeline you will see in the portal. Same vocabulary, same evidence - whether the job ran clean, looped through review, or needed an operator.
Happy path - small bug fix lands in one PR
ADO · Bug 4821 · billing-svc
An AI-enabled bug story is picked up by polling. Refinement clarifies the repro, implementation patches the idempotency key, and the PR merges after review.
Review-fix loop - reviewer requests a tighter window
ADO · User Story 5104 · checkout-api
First implementation passes CI, but the reviewer flags a replay window issue. The review comment becomes a new episode and a review-fix phase runs.
Failure and repair - provider error, explicit operator retry
ADO · Task 6620 · search-indexer
Implementation hits a provider rate limit on attempt 1. The job halts visibly. An operator inspects the attempt and triggers a retry.
Every event above is exportable. Nothing the agent does is invisible - that is the point.
Book a walkthroughBring AI work into the same process you already trust.
Connect your work-item systems, repositories, and model runtimes. Keep every automated step visible and governed.