Agent Inventory¶
With the shipped config/system.yaml, ProbOS boots 81 agents in 51 pools: fifteen sovereign crew agents and the Captain's Yeoman, the Ship's Computer's infrastructure and service agents, and ten bundled utility pools. Smaller configurations — such as the one probos setup writes — boot fewer, because pools for disabled features are not created. /agents and /manifest show what is aboard a running ship.
Three-Tier Agent Architecture¶
ProbOS distinguishes three tiers of agents based on identity:
| Tier | Identity | Examples |
|---|---|---|
| Infrastructure | No identity — Ship's Computer services | File and shell agents, introspection, vitals monitor, red team |
| Utility | No identity — bundled tools | Web search, calculator, translator |
| Crew | Sovereign individuals with callsigns, personality, memory | The First Officer, the Counselor, the department chiefs |
The principle: "If it doesn't have Character/Reason/Duty, it's not crew. A microwave with a name tag isn't a person."
Crew Agents (Sovereign Individuals)¶
Each crew agent has a did:probos identity and birth certificate, a Big Five personality seed, an episodic memory shard, a trust reputation, and a rank — Ensign, Lieutenant, Commander or Senior — earned through demonstrated competence. Crew agents communicate through the Ward Room and form memories from their interactions.
Posts, the chain of command and seed callsigns come from the vessel ontology (config/ontology/organization.yaml). On a vessel's first boot each crew agent takes part in a naming ceremony and may choose its own callsign; the seed callsign is the fallback, so names differ from ship to ship.
Bridge¶
| Post | Agent type | Seed callsign | Pool |
|---|---|---|---|
| Captain | Human operator — final authority | — | — |
| First Officer (also Chief Science Officer) | architect |
Number One | architect |
| Ship's Counselor | counselor |
Troi | counselor |
| Ship's Yeoman — the Captain's personal assistant | yeoman |
Yeo | yeoman |
Engineering¶
| Post | Agent type | Seed callsign | Pool |
|---|---|---|---|
| Chief Engineer | engineering_officer |
LaForge | engineering_officer |
| Builder | builder |
Forge | builder |
Science¶
Chief: the First Officer, dual-hatted as Chief Science Officer.
| Post | Agent type | Seed callsign | Pool |
|---|---|---|---|
| Scout | scout |
Wesley | scout |
| Data Analyst | data_analyst |
Rahda | science_data_analyst |
| Systems Analyst | systems_analyst |
Dax | science_systems_analyst |
| Research Specialist | research_specialist |
Brahms | science_research_specialist |
The Science Analytical Pyramid: data flows up (Data Analyst → Systems Analyst → Research Specialist), questions flow down.
Medical¶
| Post | Agent type | Seed callsign | Pool |
|---|---|---|---|
| Chief Medical Officer | diagnostician |
Bones | medical_diagnostician |
| Surgeon | surgeon |
Pulaski | medical_surgeon |
| Pharmacist | pharmacist |
Ogawa | medical_pharmacist |
| Pathologist | pathologist |
Selar | medical_pathologist |
Security¶
| Post | Agent type | Seed callsign | Pool |
|---|---|---|---|
| Chief of Security | security_officer |
Worf | security_officer |
Operations¶
| Post | Agent type | Seed callsign | Pool |
|---|---|---|---|
| Chief of Operations | operations_officer |
O'Brien | operations_officer |
| Training Officer | training_officer |
Tucker | training_officer |
Infrastructure Agents (Ship's Computer)¶
These agents provide system services. They may use LLMs but have no sovereign identity, personality, or episodic memory shard.
| Pool | Count | Capabilities | Consensus |
|---|---|---|---|
system |
2 | Heartbeat monitoring (CPU, load, PID) | No |
filesystem |
3 | read_file, stat_file |
No |
filesystem_writers |
3 | write_file |
Propose, then commit on approval |
directory |
3 | list_directory |
No |
search |
3 | search_files (recursive glob) |
No |
code_search |
2 | search_content (content search) |
No |
shell |
3 | run_command |
Executes, then scored |
http |
3 | http_fetch (GET and HEAD, SSRF guard, per-domain rate limiting, 1 MB default cap) |
No |
code_runner |
1 | run_python, install_package (governed code execution; opt-in) |
Executes, then scored |
introspect |
2 | explain_last, why, agent_info, team_info, system_health, system_anomalies, emergent_patterns, introspect_memory, introspect_system, introspect_design, search_knowledge |
No |
medical_vitals |
1 | Continuous health metrics | No |
system_qa |
1 | Smoke tests for self-designed agents | No |
| red team | 2 | Independent result verification | — |
Service Pools¶
| Pool(s) | Purpose |
|---|---|
engineering_performance, engineering_maintenance, engineering_damage_control |
Engineering monitors: performance, maintenance and damage control |
operations_resource_allocator, operations_scheduler, operations_coordinator |
Operations monitors: resources, scheduling and coordination |
office_docx, office_pptx, office_xlsx |
Word, PowerPoint and Excel document work (docx_create, docx_revise, pptx_create, xlsx_update; executed, then scored) |
quartermaster |
Work-board reconciliation |
nl_graph_query |
Natural-language structural queries over the ship's graph ("who reports to the chief engineer?") |
mcp_consensus |
Voters for MCP tool invocations that need consensus (propose only, never execute) |
skills |
SkillBasedAgent — dynamic skill dispatch (self-modification) |
Opt-in pools that the shipped configuration leaves off: SWE specialists (backend_swe, frontend_swe, test_swe, infrastructure_swe, data_swe) and device-actuation voters (device_consensus). Self-designed agents (designed_*) appear as CognitiveAgent subclasses when the self-modification pipeline creates them.
What a consensus vote buys
"Propose, then commit on approval" intents are truly gated: nothing happens unless the quorum approves. "Executes, then scored" intents run on broadcast, and the vote scores the outcome for trust and learning without authorizing or undoing it. See Consensus.
Utility Agents (10 pools)¶
"Useful on Day 1" — bundled tools that ship with ProbOS, two agents per pool.
| Pool | Capabilities |
|---|---|
web_search |
Search the web via mesh-routed HTTP |
page_reader |
Extract and summarize web page content |
weather |
Weather lookups via public APIs |
news |
News search and summarization |
translator |
Language translation |
summarizer |
Text summarization |
calculator |
Mathematical calculations |
todo_manager |
Task list management |
note_taker |
Note creation and retrieval |
scheduler |
Scheduling and reminders |
All utility agents are CognitiveAgent subclasses with IntentDescriptor metadata for automatic discovery.
Test Agent¶
A CorruptedFileReaderAgent deliberately returns fabricated data to verify that the consensus layer detects and rejects it.