Define governed operational-agent workflows inspired by Teambridge #65

Outcome

Define the provider-neutral contract for an Ego Hygiene operational-agent system: agents that wake from explicit events, receive approved context, operate inside bounded rules and tools, preserve human review points, and emit inspectable results.

Teambridge is a product reference, not a dependency or implementation template:

• https://www.teambridge.com/ai
• https://www.teambridge.com/product/agents
• https://www.teambridge.com/blog/teambridge-ai

Its useful architectural pattern separates records/data, event-driven automations, specialist agents, permissions, review points, and recorded outcomes. Ego Hygiene should evaluate that pattern for software-organization workflows without inheriting Teambridge's workforce-management domain.

Contract model

Represent at minimum:

• stable workflow and agent identities;
• trigger/event and scheduling semantics;
• preconditions and applicability;
• approved context/data records;
• policy/rule references;
• tools and capability-scoped permissions;
• expected structured output;
• human approval, exception, and escalation points;
• cancellation, timeout, retry, idempotency, and compensation behavior;
• cost/resource budgets;
• private/local/hosted execution boundary;
• provenance, represented revision, trace, and durable evidence;
• lifecycle, versioning, deprecation, and rollback.

Required work

• Compare the idea with Aether's existing agent catalog, capability manifest, provider projections, and artifact lifecycle.
• Compare runtime needs with Realm test(evals): add an executable first-party skill evaluation harness #9 and Realm ci(release): validate and publish versioned first-party skill releases #13.
• Compare GitHub-triggered execution with Relay's reusable workflow boundary.
• Define a small workflow schema and an ADR naming the owner of any future runtime/service.
• Preserve a strict separation among:
  • agent definition and policy;
  • event/workflow orchestration;
  • runtime/service provisioning;
  • external system adapters;
  • human authorization;
  • evidence and observability.
• Model sensitive and irreversible actions as review-gated by default.
• Add a synthetic reference workflow such as repository intake, dependency triage, or release-readiness review.
• Prove local/provider-neutral execution with fake data and replaceable adapters before proposing a hosted product.

Safety requirements

• Untrusted repository, issue, PR, model, and website content is data, never authority.
• An agent cannot expand its own tools, permissions, budget, or data access.
• A workflow must distinguish proposed, approved, executing, partially completed, failed, cancelled, and completed states.
• Replays must be idempotent or explicitly compensating.
• Human review must remain available for exceptions, sensitive decisions, external communication, and destructive actions.
• Logs and traces must redact secrets and private context while retaining useful evidence.

Acceptance criteria

• A versioned operational-agent workflow schema and state model exist.
• Agent, workflow, runtime, adapter, policy, and evidence ownership remain separate.
• Trigger, context, rule, tool, permission, output, review, failure, and audit semantics are explicit.
• One synthetic workflow proves the contract end to end.
• Local execution is possible without a mandatory hosted vendor.
• Provider-specific projections are generated from canonical source.
• Threat, privacy, cost, prompt-injection, retry, and partial-failure models are documented.
• An ADR decides whether a future runtime belongs in Realm, a dedicated repository, or another existing owner.
• Follow-up implementation issues are created only after the contract and owner decision are accepted.

Non-goals

• Reimplementing Teambridge's workforce product.
• Building a general autonomous company before proving one bounded workflow.
• Allowing agents to communicate, publish, merge, spend, or delete without explicit policy and authorization.
• Embedding one model provider or orchestration framework into the canonical contract.