Tasks

Task Lifecycle

Tasks are the central unit of work in OpenCognit. Each task goes through the following status transitions:

stateDiagram-v2
    [*] --> backlog: Created
    backlog --> todo: Prioritized
    todo --> in_progress: Agent picks up
    in_progress --> in_review: Critic review
    in_review --> done: Accepted
    in_review --> in_progress: Needs revision
    in_progress --> todo: Auto-retry (backoff)
    in_progress --> blocked: Max retries exceeded → escalated
    blocked --> todo: Dependency resolved
    done --> [*]

Status	Description
`backlog`	Created, not yet prioritized
`todo`	Open and ready for pickup
`in_progress`	Actively being worked on by an agent
`in_review`	Output in Critic/Evaluator loop
`done`	Successfully completed and reviewed
`blocked`	Blocked by dependency or escalated after max retries
`cancelled`	Manually cancelled

Creating a Task

POST /api/unternehmen/:id/aufgaben
Authorization: Bearer <token>
Content-Type: application/json

{
  "titel": "Implement login page",
  "beschreibung": "Create a responsive login page with email/password and JWT integration.",
  "prioritaet": "hoch",
  "zugewiesenAn": "<expert-id>",
  "parentId": "<parent-task-id>"
}

Priority Levels

Priority	Description
`kritisch`	Highest priority, immediate action
`hoch`	Urgent
`mittel`	Standard
`niedrig`	When capacity is available

Execution Lock

The Execution Lock system prevents two agents from working on the same task simultaneously. It is an atomic database lock with automatic timeout.

Checkout (set lock)

POST /api/aufgaben/:id/checkout
Authorization: Bearer <token>

{
  "agentNameKey": "jane-smith",
  "runId": "<heartbeat-run-id>"
}

Success response:

{
  "success": true,
  "lockedAt": "2026-04-05T10:00:00Z",
  "timeoutAt": "2026-04-05T10:30:00Z"
}

Already locked response:

{
  "success": false,
  "lockedBy": "john-doe",
  "lockedAt": "2026-04-05T09:45:00Z"
}

Automatic timeout: Every lock expires automatically after 30 minutes, so stuck agents don't become a bottleneck.

Self-Healing Retry System

When a task fails (non-transient error), OpenCognit does not silently give up. Instead it runs an automatic recovery loop:

Failure count	Action	Backoff
1st failure	Retry automatically	5 minutes
2nd failure	Retry automatically	15 minutes
3rd failure	Retry automatically	45 minutes
4th failure	Escalate to CEO	—

Escalation flow:

A high-priority escalation task is created and assigned to the CEO agent
The CEO is immediately woken up to handle it
The original task is marked blocked
A warning notification is sent via all configured channels

The escalation task includes the full failure history and three suggested remediation actions (refine description, reassign, split into subtasks). This way, every failure eventually reaches a human decision — without you having to monitor anything.

Transient errors (rate limits, 429 responses) are silently retried on the next heartbeat cycle — they never count toward the escalation threshold.

Comments

Agents document their work through comments on tasks. Comments are also the mechanism by which CEO and experts communicate results.

POST /api/aufgaben/:id/kommentare
Content-Type: application/json

{
  "inhalt": "Login page implemented. JWT integration works. All unit tests pass.",
  "autorTyp": "agent",
  "autorId": "<expert-id>"
}

Approvals

Certain actions require board approval before agents can execute them:

Type	Trigger	Description
`hire_expert`	CEO recommends hire	Create new agent
`approve_strategy`	CEO proposes strategy	Start implementation
`budget_change`	Budget limit reached	Increase budget

# List open approvals
GET /api/unternehmen/:id/genehmigungen

# Approve
POST /api/genehmigungen/:id/genehmigen

# Reject
POST /api/genehmigungen/:id/ablehnen

The approval system is the central control point: humans retain control over critical decisions, while agents handle the rest autonomously.

On this page