Case studies
How Homer is run in practice: a real setup, the problem it solves, the configuration that does it, and what it leaves open. Each one is a pattern you can copy, not a feature of Homer; the reference documents stay the authority on what Homer itself does.
Workflows and people
| Case study | In short |
|---|---|
| A ticket-to-merge-request pipeline on LangGraph | An orchestrator whose state lasts days: fan-out to worker agents, human approval gates, and saving paid work when a step hits its turn cap. |
| Keeping long runs from losing work | A per-ticket journal and work-in-progress snapshots, so a killed or re-triggered run continues instead of starting over. |
| Answering operator questions from chat | A bridge over Homer's questions API mirrors every open question into chat, and reminders nudge only in working hours. |
| A review → fix loop that converges | An arbiter, deterministic rails and a round cap that hands over to a human through ask-and-dispatch. |
| A factory that proposes its own fixes | An agent that reads run history and opens one merge request per proposed fix, without slowing the pipeline it studies. |
Security and isolation
| Case study | In short |
|---|---|
| Runs that hold no credentials | A proxy beside each run adds the tokens, so a prompt-injected run has no token to leak. |
| A public chat intake for untrusted requests | Anyone may start runs from chat: per-thread locks, global caps, a path guard and a maintainer release gate. |
| Containing a Claude step: disallowedTools, not allowedTools | With permissions skipped, allowedTools restricts nothing; disallowedTools keeps a step from spawning subagents. |
| LLM decisions that must carry proof | A model may dismiss a security finding only with a verbatim quote that a script checks. |
| An agent as an MCP tool for another team | Another team's agent calls one Homer agent over MCP, with a token that may run that agent and nothing else. |
Capacity and scheduling
| Case study | In short |
|---|---|
| Runner pools on Kubernetes | A performance pool (2 runs at a time) for builds and a light pool (8 at a time) for everything else, picked by runnerTags. |
| One lock across several agents | maxParallelRuns is per agent: a lock key shared by every agent that uses one build host. |
| One lane at a time, without parked runs blocking it | One pipeline computing at a time, where a run parked at a human gate does not hold the lane, and a watchdog for dead runs. |
| Warm caches for short runs | Cache roots on a volume and a cron agent that keeps them fresh: minutes saved per stage. |
| macOS build hosts for mobile builds | From an ssh runner behind a tunnel to self-registering pull-mode runners on macOS VMs. |
Operations
| Case study | In short |
|---|---|
| Agent changes without downtime on Kubernetes | Agents on their own volume: an agent-only change swaps the files and calls the reload endpoint, with no pod restart. |
| Linting agent definitions in CI | The checks that catch, before merge, the agent mistakes Homer would only report at runtime. |
| Accurate cost through an LLM gateway | Pricing runs from the gateway's real spend, which arrives late, plus cost by tag and cost limits. |
Have a setup worth sharing? Open an issue describing it; see CONTRIBUTING.md.