Skip to main content

How to create an agent task

Dispatch work to AI agents through Plane (project management) and Astra (agent control plane).

How agent dispatch works​

  1. Plane is the task management platform (hosted at plane.aucert.dev)
  2. Astra is the Agent Workforce Control Plane — it manages agent identities, credentials, and dispatch
  3. Agents are registered in Astra with platform accounts (Plane, GitHub, Slack)
  4. When a Plane task is assigned to an agent, Astra coordinates the dispatch

Steps​

Step 1: Create the task in Plane​

  1. Go to plane.aucert.dev
  2. Create a new issue in the appropriate project
  3. Write a clear description following the spec-driven format:
    • What: What needs to be done
    • Why: Context and motivation
    • Acceptance criteria: How to verify completion
    • Scope boundaries: What is NOT in scope

Step 2: Prepare context for the agent​

Agents work best with structured context. Include:

## Task: Add validation to KG node creation

### Context
- Module: `backend/platform/src/.../knowledgegraph/`
- Spec: `backend/platform/spec/knowledge-graph.md`
- Related proto: `proto/knowledge-graph.proto`

### Requirements
1. Validate node labels against allowed types
2. Validate edge relationships match schema
3. Return structured error responses

### Constraints
- Do NOT modify proto schemas
- Follow existing validation patterns in the codebase
- Write tests for all validation rules

Step 3: Assign to the agent​

In Plane, assign the task to the appropriate agent member. Each agent has:

  • A fun name (personality identity in Astra)
  • A Plane account (for reading/updating tasks)
  • A GitHub account (for creating PRs)
tip

Check the Agent workforce overview to see which agents are available and their specializations.

Step 4: Monitor progress​

Track agent work through:

  • Plane: Task status updates and comments
  • GitHub: PR creation and CI results
  • Astra dashboard: Agent activity and performance metrics at astra.aucert.dev

Step 5: Review the output​

Agent-created PRs follow the same workflow as human PRs:

  1. Review the code changes
  2. Verify tests pass
  3. Check context file updates are included
  4. A human must approve and merge — agents never merge their own PRs

Best practices for agent tasks​

Do:

  • Provide specific file paths and module references
  • Link to relevant spec files
  • Set clear acceptance criteria
  • Keep tasks focused (one concern per task)

Don't:

  • Assign vague tasks ("improve the codebase")
  • Skip context — agents perform better with more context
  • Expect agents to make architectural decisions — those should be in specs
  • Assign Phase 2/3 work — agents follow the same phase boundaries as humans

What's next​