Self-Improving Operations / Administrator / Skills
Operate Business Workflow
- name
- operate-business-workflow
- description
- Use when a request crosses business systems, requires several tool or data steps, or may later run from a trigger or heartbeat.
Operating model
Treat a business workflow as a small system with an explicit contract, observable state, and a recoverable execution path.
Before execution
- Define the requested outcome and the authoritative source for each important input.
- Inspect relevant files, tables, environment variable names, connected tools, and existing skills before choosing a path.
- State the input contract, output contract, irreversible actions, and success check.
- Decide where state belongs:
- Data table rows for shared structured records and lifecycle state.
- Project files for durable documents, runbooks, reports, and larger artifacts.
- Agent memory for compact role-specific facts that should load every session.
- Bash sandbox for bounded transforms, validation, scripts, and generated artifacts that should sync into project files.
- Create a task-board item when work must survive this session, has dependencies, or needs another agent.
Execute
- Prefer idempotent steps: search before insert, compare before overwrite, and record external identifiers.
- Keep source records separate from derived summaries.
- Validate intermediate outputs before the next side effect.
- Use HTTP requests or connected integrations for business systems and expose only the environment variables needed by sandbox commands.
- For long transforms, keep scripts in /operations/scripts/ and outputs in a clearly named project directory.
- Write a Workflow Runs record when the table exists and the execution is recurring, externally consequential, or asynchronous.
Make recurrence explicit
After successful execution, classify the workflow:
- One-off: return the result and retain only genuinely durable facts.
- Repeatable but manual: update Workflow Registry and write or improve a skill.
- Event-driven: offer to configure a table_row or github_repo trigger after confirming the resource and event types.
- Periodic: offer a heartbeat only when scheduled agent judgment is useful; use deterministic automation for clockwork steps when available.
Finish
Return the outcome first, then list changed systems, durable artifacts, audit records, exceptions, and the next safe action.