| name | workflow_creator_skill |
| description | Create, optimize, and evaluate reusable AgentEvolver HTML Workflows. Use when a repeated successful multi-agent orchestration should be retained, when an evolvable Workflow has recurring orchestration-level defects, or when a live version needs evidence for keep, improve, or rollback decisions. |
| version | 1.2.0 |
| type | ["orchestrator","worker"] |
| license | Apache-2.0 |
| category | meta |
| requirements | ["cpu"] |
| metadata | {} |
Workflow Creator
One methodology for the complete Workflow lifecycle. A Workflow is a reusable orchestration
program, not an Agent subtype and not a substitute for a Skill's domain instructions.
Orchestration
MetaAgent uses this lifecycle:
- Prefer an existing active Workflow when applicability and contracts match.
- Generate only after a useful orchestration pattern has repeated successfully, or when the
user explicitly requests a reusable Workflow.
- Register generated/optimized HTML as
status="active"; like Tool and Skill, it is live immediately.
- Dispatch
workflow_evaluate_agent on at least three representative cases.
- Record every outcome through
evolution_tool.record_workflow_evaluation.
- Keep a healthy version; otherwise optimize, rollback, or unload it.
Do not evolve from a single transient failure. Fix Agent/Tool/Skill quality at its owning
layer; change Workflow only for topology, routing, conditions, fan-out, verification,
retry, aggregation, budget, or termination defects.
Creating a Workflow
- Inspect active Workflows first and avoid semantic duplicates.
- Generalize concrete task details into typed
<inputs>.
- Write one complete HTML document to
extension/workflow/{name}.html.
- Use a
<workflow> element with name, semantic version, schema-version="1.1.0",
status="active", a precise description, and enable-evolving="true".
- Include
<applicability> tags and prose explaining when to use and when not to use it.
- Use only supported callable nodes:
agent, tool, skill, connector, environment,
workflow. Connector and Environment nodes require an action.
- Use only bounded control nodes:
parallel, map, branch, loop, verify, reduce,
checkpoint. Every loop requires max-rounds; all fan-out needs a concurrency bound.
- Values use restricted
${path} expressions. Never embed JavaScript, Python, handlers,
shell commands, or arbitrary templates.
- Declare complex input constraints with sibling
<schema for="name"> Draft 2020-12 JSON
Schema and keep its type consistent with the <input> attribute.
- Define explicit
<outputs> and reference guaranteed top-level step results.
- Set node
timeout and retry/backoff policy where an external capability can stall.
- Compile the file before completion and include its absolute path in
done_tool.reasoning.
Improving a Workflow
- Call
inspect_workflow first. Stop if missing or enable_evolving=false.
- Read evaluation evidence and identify a Workflow-owned defect.
- Make the smallest structural change; preserve public name and input/output compatibility
unless the task explicitly authorizes a breaking change.
- Increment the semantic version and keep status
active.
- Compile and check boundedness, reachability, capability names, and output references.
- Include the edited absolute HTML path in
done_tool.reasoning for registration.
Never tune a Workflow to one benchmark case. Prefer parameterization over copying variants.
Evaluating a Workflow
Evaluation is read-only:
- Call
inspect_workflow; confirm name, version, active status, source, and contract.
- Check schema safety, bounded termination, path coverage, capability existence, retry and
verification policy, input/output clarity, and applicability precision.
- Run a representative case when safe. Compare with the prior active version or manual
orchestration when available.
- Score outcome quality from 0.0–1.0 and report success, real run id, case id, elapsed time, token cost,
and concrete notes.
- Call
evolution_tool with action=record_workflow_evaluation. Recommend keep,
optimize, rollback, or unload from concrete evidence.
A version is provisionally healthy after at least 3 evaluations, success rate at least
0.8, and average quality at least 0.7. This summary guides retention; it does not create a
second publication state. Representative coverage and the absence of safety or termination
defects remain mandatory.