Skip to main content

Task System

Free

You don't manage these agents by chatting at them and hoping for the best. You give them tasks with clear scope, acceptance criteria, and a review gate. They brief you before they start. You approve before anything ships.

Two Systems, One Dashboard

Cohort runs two task systems side by side. Use whichever fits the work -- or both at once.

[T] Agent Tasks
Parallel. Assign work to specific agents. Each task gets its own channel where the agent asks clarifying questions, you confirm scope, and work begins. Multiple agents can work on different tasks simultaneously.
Best for: code reviews, research, content drafting, analysis -- work that needs a conversation before execution.
[Q] Work Queue
Sequential. Priority-ordered queue with a single-active constraint. Only one item runs at a time. When it finishes, the next highest-priority item claims automatically. No collisions, no race conditions.
Best for: deployments, migrations, builds, CI jobs -- work that must run one at a time in order.

How Agent Tasks Work

Every task goes through a structured lifecycle. The agent doesn't just get instructions and disappear -- it briefs you on its plan, you confirm, it executes, and the output lands in a review queue.

Briefing --> Assigned --> In Progress --> Complete --> Review
1

Briefing

You assign a task to an agent. Cohort creates a dedicated channel for that task and the agent responds with questions. What's the goal? What's in scope? What does "done" look like? This is a conversation, not a form. The agent is building a brief: goal, approach, scope, and acceptance criteria.

2

Confirmation

Once the agent has enough context, it presents a confirmation card -- a structured summary of what it understood. You read it, adjust if needed, and confirm. Now both sides agree on what "done" means before any work starts. Tasks stuck in briefing for more than 4 hours auto-fail so nothing hangs forever.

3

Execution

The agent works. Runs in a background thread with a typing indicator in the task channel so you can see it's active. The agent can use local LLM inference, CLI tools, or the Anthropic API -- whichever backend fits the work. If the agent has partnerships configured (e.g., security agent reviews all code tasks), a consultation gate fires and partners can weigh in before execution proceeds.

4

Output

The result is posted to the task channel and automatically scanned for secrets -- AWS keys, API tokens, passwords, private keys are all redacted before storage. The output includes what was produced and how it maps to the acceptance criteria from the briefing.

5

Review

Completed tasks land in a review queue. You see the output, the original brief, and record a verdict: approve, reject, or needs work. The verdict is stored with the task -- full audit trail of who asked for what, what was produced, and whether it was accepted.

The Work Queue

When work needs to run in order -- one thing at a time, highest priority first -- use the queue.

Queue state
ACTIVE Deploy staging environment
CRITICAL
Claimed 2m ago by system
2. Run integration test suite
HIGH
3. Generate weekly report (depends on #2)
MEDIUM
4. Clean up temp artifacts
LOW
Priority levels: critical > high > medium > low. FIFO within each level. Items with unmet dependencies wait regardless of priority.
1
Active at a time
Single-active constraint
4
Priority levels
critical / high / medium / low
DAG
Dependency tracking
depends_on blocks until met

Recurring Schedules

Tasks that need to run on a schedule -- daily reports, hourly checks, weekly audits. Set it once, Cohort fires it automatically.

Interval
Every N seconds. Minimum 5 minutes. "Run this health check every hour."
schedule_expr: "3600"
Cron
Standard 5-field cron expressions. "Generate the briefing at 9am on weekdays."
schedule_expr: "0 9 * * 1-5"
Once
Fire at a specific future time. "Run this migration at 2am tonight."
schedule_expr: "2026-03-20T02:00"
[!]

Auto-Disable on Failure

If a scheduled task fails 3 times in a row, the schedule auto-disables and you get a notification. No runaway failures. No silent breakage. You see it, you fix the underlying issue, you re-enable. Max 50 active schedules, with the last 100 run results kept per schedule for debugging.

Built-In Safety

Secret Scanning

Every task output is scanned before storage. AWS keys, API tokens, passwords, private keys, PII patterns -- all automatically replaced with [REDACTED:type]. Agents can't accidentally leak credentials in their output.

Stale Task Reaping

Tasks stuck in the briefing phase for over 4 hours auto-fail. No orphaned tasks sitting forever in limbo waiting for a confirmation that never comes.

Consultation Gates

If an agent has partnerships configured -- like "security agent reviews all code tasks" -- a gate fires before execution. Partners can review the plan in the task channel. The audit trail records who was consulted.

Human Review Gate

Nothing ships without your sign-off. Completed tasks land in a review queue with the original brief alongside the output. You approve, reject, or send back for rework. Full audit trail.

Full Traceability

Every task records a trigger-action-outcome triad. Six months from now, you can look at any task and know exactly what happened and why.

Task record
// How was this task initiated?
"trigger": {
  "type": "scheduled",
  "source": "sched_daily_report",
  "fired_at": "2026-03-19T09:00:00"
}

// What did the agent do?
"action": {
  "tool": "executive_briefing",
  "parameters": { "hours": 24 }
}

// What was the result?
"outcome": {
  "type": "report",
  "success_criteria": "HTML report with agent narratives",
  "artifact_ref": "briefing_2026-03-19.html",
  "verified": true
}

Use It From Anywhere

Same task system, four interfaces. Pick the one that fits how you work.

Dashboard
Visual task board with tabs for active, scheduled, completed, and archived tasks.
MCP Tools
6 MCP tools for task + queue operations. Works from Claude Code or any MCP client.
REST API
Full CRUD + claim endpoint. Build your own integrations, CI hooks, or monitoring.
VS Code
Task panel in the extension. Assign, monitor, and review without leaving your editor.

Stop Hoping AI Gets It Right

Briefing, execution, review. Every task, every time.

Get Started All Tools