Skip to Content

Task / Agent (task)

Launches a specialized sub-agent to handle a complex, multi-step task autonomously.

Parameters

ParameterRequiredDescription
descriptionYesShort (3-5 word) description for tracking
promptYesDetailed instructions for the agent
subagent_typeYesAgent type (e.g. general-purpose, Explore, coordinator)
run_in_backgroundNotrue to run concurrently without blocking
multi_sampleNotrue to enable multi-sample retry on failure

Built-in agent types

TypePurpose
general-purposeComplex multi-step tasks, code search
ExploreFast codebase search and analysis
verifyReview changes for correctness
coordinatorOrchestrate multi-agent work (spawns its own sub-agents)
planDesign implementation plans

Custom agents defined in .proto/agents/ or ~/.proto/agents/ are also available.

Background execution

Set run_in_background: true to run the agent concurrently. The main conversation continues while the agent works. A completion notification is injected at the next tool boundary.

Multi-sample retry

Set multi_sample: true for high-stakes tasks. The harness retries up to 2 more times with escalating temperatures if the first attempt fails, and returns the best result. Use for complex implementation tasks, not for searches.

How it works

  1. The sub-agent receives the prompt with its tool allowlist.
  2. It runs to completion using its available tools.
  3. The result is returned as a message to the parent agent.
  4. Sub-agents are stateless and single-use.

Sub-agents inherit tools from their parent unless an explicit tools allowlist is defined in the agent config. See Guides → Use Sub-Agents for configuration details.

Last updated on