Agent Reference Architectures

Production-shaped examples of autonomous agents built on Flue. Each example under examples/ is a complete, independently-clonable Flue app — copy one folder, set its secrets, deploy.

They share a common architecture (below); they differ in the channel that triggers them and the deploy target they run on.

Examples

Examples are named <function>-<primary-stack>. The folder name is just a unique handle — find your fit by the columns below, not the name. A customer filters by what they have: work source, code host, trigger, and where it runs.

Example	Function	Work source	Code host	Extras	Trigger	Deploy	Status
triage-jira-k8s	triage a ticket, post enriched analysis back	Jira	GitLab	Confluence	webhook	Kubernetes (EKS)	✅ complete
triage-jira-gitlab-runner	same triage, one-shot in CI	Jira	GitLab	Confluence	Jira → GitLab pipeline trigger	GitLab runners	✅ complete
triage-github-actions	triage an issue, post enriched analysis back	GitHub	GitHub	—	triage label → on: issues	GitHub Actions	✅ complete
github-pr-label-actions	act on a labelled PR; ships Dependabot auto-merge	GitHub	GitHub	—	PR label → on: pull_request_target	GitHub Actions	✅ complete
assistant-slack-daytona	do a task in a sandbox, reply in-thread	Slack	—	Daytona sandbox (per-thread)	Slack Events API webhook	any container host	✅ complete
assistant-slack-aws-daytona	full Claude-Tag platform: memory, scheduling, governance, spec-interview	Slack	—	S3 memory · Daytona/pluggable sandbox · EventBridge	Slack Events API webhook	AWS (API GW + Lambda + SQS)	✅ complete
ideate-scheduled-actions	survey the examples vs Flue's features; file agent-idea issues for the gaps	this repo	GitHub	Flue docs · @flue/*	schedule (hourly cron)	GitHub Actions	✅ complete

Want triage on a different stack (e.g. Linear + GitLab + GitLab runners)? That's a new row/folder, not a config switch — each example is pinned to one trigger and one deploy, because the trigger model (webhook vs CI one-shot) determines the deploy.

The shared architecture

Every example follows the same shape. Learn it once; each example is a variation.

• Agent (src/agents/<name>.ts) — pure wiring: model, sandbox, tools. No prose. The agent's framing lives in AGENTS.md, its procedure in a skill.
• Skills (.agents/skills/<name>/SKILL.md) — the procedure, discovered natively by Flue at runtime from the sandbox cwd. Detached from code, on their own release cycle. See docs/adding-skills.md.
• Channel (src/channels/<provider>.ts) — inbound only: verifies the trigger and dispatches to the agent. (CI-driven examples use a one-shot flue run instead of a long-running channel.)
• Tools (src/tools/<provider>/) — outbound only: the API calls the agent makes, grouped by provider.
• Inbound = channel, outbound = tools. A provider you receive from is a channel; one you only call is a set of tools.

See AGENTS.md for the full conventions.

Adding your own skills

You do not modify agent code to change what an agent does — you edit or add a skill. Full guide: docs/adding-skills.md.

Getting started

• New to this? Build one from scratch: docs/creating-an-agent.md.
• Have a stack in mind? Pick the matching example from the table and follow its README:

cd examples/triage-jira-k8s
cat README.md