Skip to content

WACP — Workspace Agent Coordination Protocol

WACP (Workspace Agent Coordination Protocol) defines the rules by which autonomous agents coordinate work. It is a coordination protocol — it specifies what agents say, how they communicate, and the structures they operate within.

Without a coordination protocol, multi-agent systems face:

  • Opacity — you cannot see what agents are doing or why they failed
  • Lost work — when an agent fails, you don’t know what artifacts it produced
  • Blame confusion — who is responsible for an outcome?
  • No recovery — if the system crashes, you lose all state
  • Silent failures — agents may deadlock or produce garbage without anyone knowing
  • No human control — humans cannot intervene without redesigning the system

The protocol answers five questions:

  1. Where does an agent work?Workspaces: isolated, bounded contexts
  2. How do agents communicate?Envelopes and signals: structured messages and typed notifications
  3. How is progress recorded?Checkpoints: immutable snapshots of work
  4. How is work organized?Tasks: units of work forming dependency graphs
  5. What happened?The trail: an append-only audit log

The unit of isolation. A bounded context assigned to exactly one agent. Workspaces have defined inputs (directives) and outputs (checkpoints). They form containment trees — a parent workspace creates child workspaces.

Envelopes carry content between workspaces. Signals are lightweight, typed notifications declaring state changes (ready, blocked, complete, failed).

Immutable progress records that survive agent failure. Chained together to form a complete history of work.

Structured units of work in a directed acyclic graph (DAG). Dependencies are explicit — a task cannot start until its predecessors complete.

The append-only audit log. The single source of truth for everything that has occurred. Enables recovery, observability, and security audits.

Explicit gates and escalations for human oversight. Humans can approve, inject, or redirect work at any moment. Supervision is part of the protocol, not optional middleware.

What an agent is permitted to do. Assigned at creation, immutable. Enforced by the runtime, not by prompts.

The specification comprises 20 constituent specs organized into four layers:

LayerSpecsPurpose
FoundationsClock, RolesTime ordering and permission model
PrimitivesWorkspace, Envelope, Signal, Checkpoint, Trail, Task, Identity, UserCore protocol objects
TopologyGraph, Tree, Visibility, Ownership, Causation, ChannelsStructural relationships
MechanismsIntegration, Security, Human-Highway, RecoveryCross-cutting operations
  • Messages over mutations — agents never modify shared state directly
  • Roles are structural, not suggested — permissions are walls, not guidelines
  • Explicit lifecycle — every state transition is declared
  • Context is scoped, not shared — least privilege applied to attention
  • History is first-class — the trail is write-ahead, immutable, hash-chained
  • Human control is first-class — gates fire at specific points, escalations activate on blocks

Draft v0.1 — architecturally complete, all 20 specs drafted.