Continuity
Homes, identity and unfinished work survive disconnects, restarts and lost Pods.
Design principle
the open-source company harness.
Build autonomous companies, under human control.
Turn persistent Agents into an accountable organization with durable work, explicit authority and a verifiable path from human intent to delivered evidence.
Pi harnesses a model into an Agent. AgentOS harnesses Agents into a company—without replacing the repositories, trackers, infrastructure or human decisions you already trust.
One prompt
Read https://github.com/akua-dev/agentos/blob/main/BOOTSTRAP.md.
Help me bring AgentOS online — check my setup first, ask before changing anything.The coding Agent you already use
Read https://github.com/akua-dev/agentos/blob/main/BOOTSTRAP.md.
Help me bring AgentOS online — check my setup first, ask before changing anything.◇ Environment inspected◆ Changes wait for your approval● First Mate readyWork can start wherever thinking is fastest.
Start anywhere. Bring in the Fleet when the outcome matters.
Shape a UI, sketch backend architecture, research, implement, diagnose, test, or conduct a code review locally. Package the desired result and durable evidence, then let First Mate turn accepted intent into accountable Fleet work.
1
Working evidence
A UI direction, architecture sketch, research finding, patch or review.
2
Bounded handoff
Outcome, evidence, constraints and exact authority boundary.
3
Accountable Fleet work
A Task and one active Assignment after acceptance.
4
Reviewed result
Delivery evidence returns for your next decision or iteration.
AgentOS × Flect
One front door to the work of your company.
Flect /flekt/ — from Latin flectere: to bend, curve, or turn.
Flect makes that idea software. Its running interface bends around the person, task, and decision: shape a working UI from inside the product, keep or reject it safely, then use it with a dedicated App Agent. Install it today while AgentOS keeps ownership, decisions, and delivery durable.
Available today
The company loop
Bring the right work to the right people.
AgentOS assembles product work, personalized context and decisions, then brings in the people with relevant context, expertise, or authority. Flect's built-in Agent helps them inspect or revise the draft before bounded, explicitly approved intent returns to First Mate.
Proof before promises.
AgentOS publishes benchmark attempts—including failures—and links each result to sanitized evidence. The question is not whether one demo looked autonomous. It is how many verified outcomes the organization delivered for the human attention it used.
Responsibility
Every accepted outcome has one accountable Assignment owner and a path back to people.
Design principle
Visibility
Inspect real terminals, durable state and delivery artifacts—not a green status badge.
Design principle
Learning
Evidence becomes reviewed improvement; memory remains fallible context, never authority.
Design principle
Continuity
Homes, identity and unfinished work survive disconnects, restarts and lost Pods.
Design principle
Responsibility
Every accepted outcome has one accountable Assignment owner and a path back to people.
Design principle
Visibility
Inspect real terminals, durable state and delivery artifacts—not a green status badge.
Design principle
Learning
Evidence becomes reviewed improvement; memory remains fallible context, never authority.
Design principle
Continuity
Homes, identity and unfinished work survive disconnects, restarts and lost Pods.
Design principle
Responsibility
Every accepted outcome has one accountable Assignment owner and a path back to people.
Design principle
Visibility
Inspect real terminals, durable state and delivery artifacts—not a green status badge.
Design principle
Learning
Evidence becomes reviewed improvement; memory remains fallible context, never authority.
Design principle
Continuity
Homes, identity and unfinished work survive disconnects, restarts and lost Pods.
Design principle
Responsibility
Every accepted outcome has one accountable Assignment owner and a path back to people.
Design principle
Visibility
Inspect real terminals, durable state and delivery artifacts—not a green status badge.
Design principle
Learning
Evidence becomes reviewed improvement; memory remains fallible context, never authority.
Design principle
From one answer to a company.
Models generate. Agents own bounded outcomes. AgentOS gives many Agents the durable organization required to work as a company.
Model → generation
Chat → turns + context
Copilot → retrieval + instructions
Human still carries the goal.Assistance is not delegation.
A model generates. Chat makes generation conversational. A copilot enters the workflow with retrieval and integrations. Each can be enormously useful, but the human still preserves continuity and decides every next step.
Bring AgentOS onlinebounded outcome
→ reason
→ act with tools
→ observe reality
→ adjust
→ return evidenceAn Agent owns a bounded outcome.
The delegation threshold arrives when the system can carry a goal through a reason–act–observe loop. A harness supplies its session, context, Skills, memory, sandbox, permissions and recovery.
Give the Fleet an outcomeAgent team + durable organization
identity · ownership · hierarchy
authority · attention · recovery
= autonomous companyParallel Agents are not yet a company.
Routing sessions can increase throughput. AgentOS adds what makes that activity an accountable organization: persistent identity, durable work, explicit authority, supervision, human attention and recovery.
Add a durable domainOutcome first
The plan emerges from evidence.
You define what should become true, how it will be proven and where authority stops. First Mate reads current reality, keeps deliberate backlog durable and unassigned, gives accepted asynchronous work one accountable Assignment and returns the decisions that still belong to you.
human outcome + authority
→ inspect current reality
├─ actionable → Assignment
├─ uncertain → Scout
├─ consequential → Captain decision
└─ prerequisite → dependency
→ evidence
→ next distinct outcomePredeclared workflow
The human maintains the graph.
Expected steps become nodes. Unexpected work escapes the model, orchestration owns a shadow of state and the human watches execution.
AgentOS
The organization maintains responsibility.
Agents reason from authoritative reality. New evidence can become owned work while native systems keep the facts and people keep consequential authority.
Native systems. One accountable organization.
Tracker
Human planning and provider workflow.
PostgreSQL
Durable Tasks, Assignments and decisions.
Kubernetes
Live workloads and persistent homes.
Herdr
Inspectable native harness sessions.
Git
Delivered source and review artifacts.
Connect authorities. Do not mirror them.
AgentOS coordinates across native systems without hiding them behind a universal wrapper or shadow database. The authority for each consequence stays where it belongs.
Human intentTracker
Accountable ownerPostgreSQL
Running workKubernetes + Herdr
Delivered resultGit + provider
A crew, not a swarm.
Captain is a role a person or team can hold. First Mate turns its intent into owned work. Second Mates lead durable domains. Crewmates deliver bounded outcomes. Humans use one company surface; internal routing stays the company's problem.
Captain → First Mate → Second Mate → Crewmate
one direct supervision edge · one accountable owner
one direct supervision edge · one accountable owner
Intent has a chain of custody.
A Task preserves accepted intent. An Assignment binds the active outcome to one accountable Agent. Handoffs keep history. Evidence closes the loop; a status label alone does not.
Human intent
→ Task
→ Assignment
→ accountable Agent
→ verified resultStart with one real project.
Your repository, issues, board and CI stay where they are. The Fleet delivers ordinary pull requests through the workflow that project already trusts. Authority grows only through standing rules people deliberately record.
Adopt a projectEarly, open and inspectable.
AgentOS is early. It does not ask one successful Fleet to stand in for general production maturity. Every release names its support boundary, and every improvement is expected to begin with evidence.
How evaluation worksStop being the loop. Stay at the helm.
Bring AgentOS online, give the Fleet a real outcome and learn each new layer as your company starts to work.