Tasks, Chains, and Reviews¶
Arc supports both simple queued tasks and multi-step chains.
Single Task¶
A single task is one unit of work assigned to one task agent.
Example:
This is the simplest background job model.
Chain¶
A chain is one task that moves through multiple task agents in sequence.
Example mental model:
- step 1:
researcher - step 2:
writer - step 3:
revieweror human review
Arc stores the chain as one task with multiple steps.
Dependency Between Separate Tasks¶
Arc also supports separate tasks that wait for earlier tasks.
arc task add "Research AI frameworks" --assign researcher
arc task add "Compare the top 3" --assign researcher --after t-001
arc task add "Write recommendation" --assign writer --after t-002
--after means the later task depends on the earlier task finishing first.
That is different from a multi-step chain:
--afterlinks separate tasks- a chain keeps multiple steps inside one task
How Handoff Works Inside A Chain¶
When one step finishes, Arc passes the earlier task activity forward as context.
The next agent sees:
- the main task title and instruction
- a note that earlier steps are complete
- previous task activity and comments
- any review feedback that needs to be addressed
The current step prompt explicitly tells the next agent to continue from previous work and not repeat earlier work unless necessary.
Reviews¶
A step can optionally be reviewed by:
- another task agent
- a human
Possible outcomes:
- approve: move to the next step
- needs revision: send work back to the agent who just did that step
Bounce Limits¶
Each task has max_bounces.
In a multi-step chain, the bounce counter is effectively used per review stage, not across the whole chain, because Arc resets it when moving to the next step.
Important Current Limitation¶
A multi-step chain currently shares one overall task instruction.
The step definition itself stores:
- which agent should handle the step
- which reviewer, if any, should review it
It does not currently store a separate instruction per step.
So if you want strong step-by-step guidance today, put that guidance into the main task description in a clear structured way.