CONTROL PLANE
Queen, Director, Planner, Constraint — orchestration, planning, and policy services. They coordinate missions but are never mission workers themselves.
THE COLONY
ONE MISSION.
MANY MINDS.
The Queen holds the objective. Twelve specialized mission agents research, inspect, propose, test, review, verify, report, and remember — each inside a versioned execution contract the runtime enforces before any task runs. Not everything in the colony is an LLM worker, and no role can apply its own changes.
WHAT IS, AND ISN'T, AN AGENT
MISSION AGENTS
The twelve roles below — real executors with runtime handlers and execution contracts. The runtime rejects any task outside a role's contract.
DETERMINISTIC SERVICES
Inventory, Network Scout, Health, Proxmox, Storage, Backup, Security Scout, Change Archivist, Quartermaster — plain services, never LLM-directed. Mission agents may consume their structured data.
Anything unclassified fails closed as a visual scaffold: displayed, but not executable. Source: docs/ANT_EXECUTION.md.
THE TWELVE MISSION ROLES
All twelve roles are enabled by default on a new installation (roster_profile: full since v0.3.8.41), with per-role kill switches. Roles run only when a mission needs them. Expand a role for its contract boundary.
ResearcherGathers local context near mission intake: what the workspace contains and what the system looks like.
- Trigger
- Planner-selectable, near intake
- Accepts
- Mission objective, workspace
- Tool boundary
- system_info · list_directory · search_workspace · repository_index
- Produces
- context_brief (currently prose)
- Limits
- Read-only tools
WebExternal research when a mission needs information from outside the workspace.
- Trigger
- Planner-selectable, when external info is needed
- Accepts
- Research question
- Tool boundary
- web_search
- Produces
- source_set (typed)
- Limits
- Read-only; web access is off by default
FileInspects files inside the workspace boundary, discovering paths rather than only reading named ones.
- Trigger
- Planner-selectable
- Accepts
- Paths or search terms
- Tool boundary
- list_directory · read_text_file · search_workspace · repository_index
- Produces
- file_set (typed)
- Limits
- Read-only, confined to the workspace boundary
UI CartographerMaps user-interface structure before code changes touch it; hands its map to the Coder.
- Trigger
- Planner-selectable (gated; on in the full profile)
- Accepts
- UI source files
- Tool boundary
- 4 read tools
- Produces
- ui_map
- Limits
- Read-only; not yet mandatory before Coder
CoderTurns context into a proposed patch set. Deliberately has no tools: proposed code is data until the runtime materializes it.
- Trigger
- Planner-selectable
- Accepts
- Context from earlier roles
- Tool boundary
- Model only — no tools, deliberately
- Produces
- patch_set (proposal)
- Limits
- Proposals only. Cannot apply changes
TesterRuns allowlisted, manifest-driven checks and records reproducible evidence.
- Trigger
- Inserted by policy after every state-changing patch set
- Accepts
- Patch set, check manifest
- Tool boundary
- run_allowlisted_check only
- Produces
- test_report + evidence
- Limits
- No shell, no model; evidence required
SoldierSecurity review of the actual proposed patch content through a deterministic policy scan.
- Trigger
- Inserted by policy on every state-changing patch set
- Accepts
- Patch set content
- Tool boundary
- None — PolicyScan is an in-process deterministic service
- Produces
- security_review
- Limits
- Blocks are not model-overridable
VerifierReads stored evidence and decides whether the mission outcome is actually verified.
- Trigger
- Planner-selectable; policy binds it to Tester and Soldier evidence and inserts one when the plan omits it
- Accepts
- Evidence store
- Tool boundary
- None — reads evidence, not prose
- Produces
- verification_bundle
- Limits
- A deterministic failure cannot be overridden by model text
MedicDiagnoses retryable failures and recommends a bounded repair before work returns to the Coder.
- Trigger
- Triggered by retryable failure only
- Accepts
- Failure context
- Tool boundary
- None — consumes failure context
- Produces
- failure_diagnosis · repair_recommendation
- Limits
- 2 diagnoses per mission; repeats escalate
BuilderAssembles the operator-facing answer after verification.
- Trigger
- Planner-selectable, after verification
- Accepts
- Verified mission results
- Tool boundary
- Model only
- Produces
- operator_summary
- Limits
- Reports; does not modify anything
ScribeDrafts release notes and documentation changes from what actually changed.
- Trigger
- Planner-selectable, after verification
- Accepts
- Changed-files summary
- Tool boundary
- read_changed_files_summary
- Produces
- release_notes · docs_draft
- Limits
- Documentation-path patches only
ArchivistConsolidates what the mission taught the colony after the canonical evaluation persists.
- Trigger
- Runs after finalization, outside the task graph
- Accepts
- Canonical mission evaluation
- Tool boundary
- None — emits memory candidates
- Produces
- memory_candidate
- Limits
- Positive reinforcement only from completed_verified outcomes
No mission role has apply permission. apply_patch, shell_command, and write_text_file are structurally forbidden to every mission agent; patch application lives only in the Queen/Director approval pipeline. Tool calls are authorized before running, and denials land on the audit stream. Sources: ANT_EXECUTION.md and PLAN.md §3 at v0.3.8.67.