Automatos·ai
Build · 1.36.04 Tag · Mission Engine Status · Operational

An operating system
for autonomous agent teams.

Automatos is the open platform for AI workforces — design specialised agents, equip them with skills and knowledge, schedule their work, and run the whole room from one command centre.

Built on · the frontier · and the open
OpenRouter/ Anthropic/ OpenAI/ Google/ Llama
M.01 — Marketplace
100+
Pre-built agents, ready to install from the community marketplace.
M.02 — Tooling
1,000+
Tool integrations — GitHub, Slack, Stripe, Shopify, Notion, Datadog…
M.03 — Models
300+
LLMs through one router. Bring your own keys; mix frontier with cheap.
M.04 — Skills
150+
Reusable, versioned capability packs that drop onto any agent.
M.05 — Playbooks
80+
Multi-step playbooks — scheduled, triggered, hand-offs between agents. Drop-in or fork.
Plate / 003 · System
The system,
end‑to‑end.
Fig. 03 — Router · Mission · Memory Open · Apache 2.0 · v1.36.04
Router 7 tiers

Every incoming message walks the same ladder — fast deterministic rules first, semantic and LLM only when they earn it.

T0Overrideuser
T1Cacheredis · ~5ms
T2aRulessource · pattern
T2bTriggersjira · webhook
T2.5Semanticcosine · 0.95
T2cIntentkeywords
T3LLMclassify
Mission engine DAG · 5s tick

Goals decompose into a task graph. A stateless coordinator reconciles state every 5 seconds — plan, dispatch, verify, repeat.

/01Decomposeplanner
/02Approvehuman · hitl
/03Dispatchparallel · locked
/04Executeagent · tools
/05Verifycross-model
/06Reconcilestall · retry
/07Reportevents · cost
Memory 5 layers

A biologically‑inspired hierarchy. Sessions promote into facts; facts decay until they earn a place in long-term memory.

L0Focuscontext window
L1Workingredis · 24h
L2Short termpostgres · decay
L3Long termmem0 · qdrant
L4Knowledgerag · nl2sql
·Field memoryqdrant · shared
·Daily logstemporal
Plate / 004 From the docs
Three pages from /docs · the actual spec, not the marketing
↗  github.com/AutomatosAI/automatos-ai
Fig. 05 — Router Doc / 01

Decisions, in tiers.

The router escalates progressively. A cache hit returns in milliseconds; an unfamiliar request walks all the way to LLM classification — and the result is cached for next time.

RequestEnvelope
T0override · user
T1cache · redis
T2rules · triggers · intent
T2.5semantic · cosine ≥ 0.95
T3llm classify
RoutingDecision · logged
corrections feed cache learning

"Optimal routing accuracy while minimising latency and cost."

/docs/universal-router/_index.md read ↗
Fig. 06 — Missions Doc / 02

Goals into a graph.

A natural-language goal becomes a directed acyclic graph of tasks. State lives in the database; the coordinator is stateless and ticks every five seconds.

user goal
/01planner.decompose()
/02orchestration_run · awaiting approval
/03human · approve
/04dispatcher · optimistic lock
/05agent · execute
/06verifier · cross‑model
events · budget · telemetry

"DB-authoritative; the coordinator reconciles state on every tick."

/docs/missions-multi-agent-coordination/_index.md read ↗
Fig. 07 — Memory Doc / 03

Five layers of recall.

Sessions live in Redis; short-term memory decays on an Ebbinghaus curve; whatever survives is promoted into long-term semantic facts — searchable, scoped to a workspace.

L0focus · context window
L1session · redis · 24h ttl
consolidate on expiry
L2short‑term · postgres
·decay = importance · e^(−rate · hours)
access‑count ≥ 3 · promote
L3long‑term · mem0 + qdrant
L4knowledge · rag · tools

"Sessions promote into facts; facts decay until they earn a place."

/docs/memory-system/_index.md read ↗
Chapter 01 The workforce
Build, install & orchestrate specialised agents
↗  Marketplace · 100+ agents

Hire the agents.
Equip them with skills.
Let them work.

100+ marketplace agents — Code Reviewer, QA Engineer, Sentinel, Scribe, researcher and marketer roles, Shopify specialists, more. Each one carries its own model config, persona, capabilities and performance history. Install from the marketplace, or build a custom agent in seconds.

/01 Skills, not prompts.150+ portable capability packs — system prompt, tool set, output contract — drop onto any agent.
/02 Workspace bundles.Install a whole team in one click — agents, skills, playbooks and dashboard widgets pre-wired together.
/03 Persona & metrics.Each agent ships with a persona, model config, and visible performance over time.
/04 Build your own.From a blank canvas or by forking marketplace agents. Versioned, exportable, testable.
Chapter 02 Orchestration
The Universal Router & mission engine
↗  Architecture · v1.36

One message in.
The right agent takes it.

A four-tier router — cache, rules, semantic, LLM — sends every message to the agent best positioned to handle it. Missions chain agents into multi-step workflows with scheduling, triggers and inter-agent coordination. Sandboxed workspaces let them run code, manage files, and touch real Git repos.

/01 Universal Router.Cache → rules → semantic → LLM. Right agent, every time.
/02 Missions & Playbooks.Multi-step automations with schedules, triggers and hand-offs.
/03 Workspace execution.Sandboxed environments — run code, manage files, edit Git.
/04 Multi-tenancy.Each team isolates their agents, data and config. Default on.
Chapter 03 Visibility
The command centre, knowledge & cost
↗  Single / pane
/.01Command centre

See your whole workforce at a glance.

Live agent status, scheduled routines, completion metrics, agent reports. Auto — the operator — surfaces what needs attention, so you don't have to chase your team.

Open command centre
/.02Knowledge

Knowledge bases
with cloud sync.

Upload documents, sync folders from Dropbox or S3, let the platform chunk, embed and index automatically. Agents get RAG-powered access to everything you've ever written.

How it indexes
/.03Cost

Know what
the workforce costs.

Track every API call across every model. Per-agent and per-request spend, usage trends, projections. No surprise bills — route cheap for heartbeat, frontier for reasoning.

Cost analytics
Field note / 01

“The closest thing I've used to hiring a team — except the team is awake at 3am, knows our codebase, and reports back with what it shipped.”

K. Hollander Head of Ops · Ledger House
Chapter 04 Foundation
Open source · self-hostable · audit-friendly
↗  github.com/AutomatosAI

An open stack —
not a black box.

Apache 2.0. The whole platform — orchestrator, frontend, services — ships under a license you can fork, audit, and run on your own infrastructure. Next.js · FastAPI · Postgres · S3 · OpenRouter. No vendor wall.

/01 Apache 2.0.Fork it. Ship a private build. Contribute back when ready.
/02 Self-host.Docker-compose up on your own infra in under five minutes.
/03 BYO keys.OpenAI, Anthropic, Google, DeepSeek, Mistral, Qwen, Llama.
/04 Auditable.Every routing decision, every tool call, every prompt — in trace.
Get started · sign up

Hire your first agent
in ten minutes.

Pick a role, plug in your channels — Slack, Telegram, Gmail, GitHub — give it a knowledge base, and put it on a schedule. The first run is on us.

OpenRouter‑native 20+ channels via Composio Self-host or hosted No card required
Recently / shipped
2026·05·02
Mission scheduler v2 — cron-style triggers, fan-out hand-offs.
PRD 142. Read the post-mortem on what changed under the hood.
2026·04·19
Shopify pack — 12 agents, 32 skills, store-ops widgets.
Install once; workspace turns into a running e-commerce back office.
2026·04·06
Cost projections — per agent, per request, per route.
Frontier vs cheap routing visualised against historic load.