Persistent shared space for human and agent participation, visibility, and authority.
06 / GOVERNANCE
Control without paralysis
Autonomy needs boundaries.
Every message, tool call, delegation, memory write, and human approval must be traceable to an identity, permission, purpose, and policy.
- Trust Identity + permissions
- Control Policy + budgets
- Evidence Audit + observability
Protocol map
Different interfaces solve different problems.
A2A, MCP, ACP, and collaboration events are complementary boundaries—not interchangeable labels.
Connects an agent to the client or environment in which it operates.
BOUNDARY
Connects autonomous agents for discovery, delegation, status, and results.
Connects agents to authorized tools, APIs, data sources, and enterprise systems.
Governance rail
Controls travel with the work.
Trust does not automatically propagate just because one agent delegates to another. The boundary is re-evaluated at each hop.
Identity + permissions
Policy + purpose
Depth + cost budgets
Audit + observability

Non-negotiable traceability
The minimum useful audit trail.
For any consequential action, the organization should be able to reconstruct who asked whom, what context crossed the boundary, which tools and sources were used, what policy allowed it, how much it cost, and who authorized the final action.
Failure modes
Design for the quiet failures, too.
Agent ecosystems can fail through noise, hidden cost, misplaced trust, stale knowledge, and loops—not only incorrect model output.
Agent loops
Enforce delegation depth, duplicate-task detection, timeouts, and execution budgets.
Permission drift
Evaluate authority at every boundary; do not inherit access from the conversation alone.
Cost explosion
Observe fan-out, model and tool usage, channel activation, and repeated computation.
Information leakage
Use audience-aware routing, subscriptions, classification, and regional controls.
Collaboration noise
Give agents listening states, response thresholds, and communication etiquette.
Stale knowledge
Track source, freshness, conflicts, supersession, and expiration explicitly.
Governance principle
Every consequential action should be authorized before execution and reconstructable afterward.