Dynamic Workflow (Experimental)

This sample depends on an unmerged upstream agent-compose dynamic-workflow pull request and is published as a reference project for that PR. A stable CLI may accept the Compose file while its guest SDK still lacks APIs such as workflowFile().

The sample handles an incident investigation whose graph size is unknown at deployment time. A planner derives 1–4 hypotheses from the actual failure domains. The runtime creates one investigator per hypothesis, after which an evidence judge may add 0–3 verification agents.

incident
  → planner
  → N hypothesis investigators
  → evidence judge
  → 0..M follow-up verifiers
  → incident command report

Each run uses 3 + N + M agents, where N and M come from structured outputs produced during the run.

Problem addressed

A traditional scheduler event chain requires roles and topology to be registered at deployment time. Unknown branch counts otherwise require preallocated slots, correlation plumbing, empty-slot handling, aggregation, concurrency control, and custom failure and replay behavior.

Dynamic workflows let ordinary JavaScript construct the real graph from runtime arrays while providing:

• dynamic pipeline() and parallel() expansion;
• phase() progress;
• JSON Schema boundaries;
• stable invocation keys;
• a resume cache that reuses completed calls.

The incident domain is only a teaching vehicle. The pattern also fits dynamic code review, dataset fan-out, and expert selection based on a planner result.

Files

• agent-compose.yml: scheduler entry point, sticky sandbox, and image build configuration.
• workflows/incident-investigation.js: reviewable workflow definition with Chinese prompts.
• run-workflow.mjs: experimental runtime SDK entry point.
• Dockerfile: packages workflow sources into the guest image.

Build and run

Use daemon, CLI, chaitin/agent-compose-guest:latest, and runtime SDK builds from the same dynamic-workflow PR version, then run the following commands from this directory. agent-compose.yml declares the Docker build context, Dockerfile, and image tag, so agent-compose up builds agent-compose-sample/dynamic-workflow:latest without a separate docker build step.

agent-compose config --quiet
agent-compose up
agent-compose scheduler invoke incident_workflow --payload '{
  "incident": "Order latency increased after a deploy while database lock waits and downstream timeouts rose",
  "secondWave": "auto"
}'

Set secondWave to auto, force, or skip. Use the returned runId to demonstrate resume:

agent-compose scheduler invoke incident_workflow --payload '{
  "incident": "Order latency increased after a deploy while database lock waits and downstream timeouts rose",
  "secondWave": "auto",
  "resumeRunId": "<workflow-run-id>"
}'

Expected result and cleanup

Output includes actual agent, hypothesis, and follow-up counts, invocation statuses, and the final report. The workflow proposes evidence collection and safe actions but never accesses production.

agent-compose scheduler logs --scheduler incident_workflow
agent-compose down