Before this change, agents communicated directly with each other by
writing to a shared file (Meta/agent-messages.md), following the
protocol in references/inter-agent-messaging.md. This was fragile
and hard to control.
Now the dispatcher (CLAUDE.md) is the only coordinator. Agents do
not talk to each other. Instead, when an agent detects work for
another agent (e.g. missing vault structure, orphan notes), it
outputs a "Suggested next agent" section. The dispatcher reads it,
validates it against the agents-registry, and decides whether to
chain the next agent.
New files added:
- references/agent-orchestration.md: the new coordination protocol,
including call chain tracking, anti-recursion rules, and a max
depth of 3 agents per user request
- references/agents-registry.md: a capability registry that maps
each agent to its inputs, outputs, and chaining suggestions
Deleted:
- references/inter-agent-messaging.md: the old peer-to-peer protocol
The skills/ folder contained 8 SKILL.md files (one per agent) that
were auto-generated wrappers. They duplicated the same logic already
present in the agent files under agents/.
Claude Code discovers agents natively from .claude/agents/ by reading
their YAML frontmatter (name, description, tools, model), so the
skills layer was redundant and has been removed.
gnekt