Concepts

Meet the agents

Who the AI agents are, what each one does, and where you see them on the pipeline detail page.

Short answer

Bilbis runs your task through a team of AI agents. Each agent has one job. You don't pick the agents - they're called in the right order automatically. You'll see their names on the pipeline timeline and on questions they ask you.

How it works

A pipeline is handed off between agents the way work is handed off between people on a real team. The Work Planner figures out the overall plan. The Repo Router reads your task and picks the right repository. The Implementation Agent writes the code. The Test Runner runs the tests. The Review Agent reads the diff. The PR Publisher merges the result.

You don't control the handoffs. You just see them happen on the Timeline tab of the pipeline detail page - each row says which agent moved the pipeline.

The team

AgentWhat it doesWhen you see it
Work PlannerOwns the overall plan and unblocks the rest of the team. Steps in when the pipeline needs a higher-level decision. Also runs a Planner pass that decomposes the task into file-level steps before coding starts.Pipeline start, the Planning state, escalations.
Repo RouterReads your task, picks the right repo when several are in scope, and decides how to split the work across repos.The Routing state. Multi-repo split decisions.
Implementation AgentWrites the code. This is the agent that does the heavy lifting on most pipelines.The Coding and Addressing review states.
Test RunnerRuns the project's tests and the CI checks. Reports back what passed and what didn't.The Testing and CI states.
Review AgentReads the diff before merge and writes review comments.The Reviewing state and the Learnings tab.
PR PublisherHandles pushing, merging, and the final cleanup once approval lands.The Pushing and Merging states.
Runtime WorkerHandles infrastructure-flavored work - environment setup, build configuration.Pipelines that touch infra; less common.

The names map to roles, not real people. The same Implementation Agent handles every pipeline's coding work in your organization.

Where you see them

On the timeline

The Timeline tab on the pipeline detail page has one row per state transition or agent action. Each row tells you which agent took the action and when. A row that reads "Implementation Agent pushed branch feat/health-endpoint" is exactly what it sounds like.

On clarification questions

When an agent has a question, the Clarifications tab shows the question with a small badge naming the asker - Implementation Agent, Review Agent, Test Runner, and so on. See Clarifications and approvals.

In LLM calls

The LLM calls tab shows every prompt the agents made. The agent is shown next to the model and cost so you can tell which role spent how much.

In learnings

The Learnings tab on a finished pipeline shows what the Review Agent extracted as reusable knowledge - conventions, patterns, CI fixes. Those go into the Insights → Learnings library and inform future pipelines.

Why agents have names at all

Each agent has its own prompt, its own job, and its own seat in the workflow. Naming them - Repo Router, Implementation Agent, Review Agent - means timelines and questions read like a normal team conversation instead of a wall of opaque function calls. When something goes wrong, you can trace it to a specific role: "the Review Agent flagged this", "the Implementation Agent pushed the wrong branch".

Internal names you might see in logs

The agents have internal role names (Tech Lead, Senior Engineer, etc.) that operators occasionally encounter in raw telemetry, exports, or older pipeline rows. The customer-facing UI maps them to the public labels above automatically - this section is only useful when reading raw logs.

Internal rolePublic label
Tech LeadWork Planner
Project LeadRepo Router
Senior EngineerImplementation Agent
Code ReviewerReview Agent
QA EngineerTest Runner
Release ManagerPR Publisher
DevOps EngineerRuntime Worker

Even older rows carry the original Greek-mythology codenames (Zeus, Hephaestus, Athena, …). See Agents runtime names for the full decoder.

On this page