Skip to main content

Skills

Skills are purpose-built sub-agents attached to a parent Agent. They let you separate what an agent is (its identity, persona, guardrails) from what it does (specific task workflows) — instead of bundling everything into one base instruction.

Skills are unlocked by selecting Harness as the agent's orchestration mode.

Orchestration modes​

The Orchestration Mode selector lives on the Agent Builder edit screen, between Environment Variables and the Model selector. Three modes are available (visually styled like the Voice Agent execution-mode selector):

  • ReAct — the classic reasoning + tool calling loop.
  • Planning — uses the platform's planning capability, with a mid-execution evaluator and next_goal narration.
  • Harness — delegates structured tasks to Skills (sub-agents).

Mode is part of the agent's saved configuration.

Harness mode — the Agent Profile and Skills section​

When you switch an Agent to Harness mode, two changes happen on the edit screen:

  • Base Instructions is renamed to Agent Profile. Only the label changes — the field works the same way. Use it for identity, persona, and global guardrails that apply to every interaction.
  • A new Skills section appears directly below the Agent Profile.

The Skills section lists every skill attached to the agent. Each row shows the skill's title and description, with two actions:

  • Edit — opens the skill drawer pre-filled.
  • − (Remove) — deletes the skill from the agent.

A + Add button in the section header opens the skill creation drawer.

Creating a Skill​

The side drawer has four fields:

  • Title — short name for the skill.
  • Description — a plain-language description of when the skill applies. This is what the agent uses at runtime to decide whether to invoke this skill — write it like you'd write a routing instruction.
  • Skill Text — the full instructions that form the sub-agent's base instructions when the skill executes. Treat it as the system prompt for a focused sub-agent.
  • Skill Execution Type — pick one of:
    • In context — execute the skill within the parent agent's context.
    • Sub Agent — execute as a separate sub-agent run.

Click Create to save. The drawer closes; the skill appears in the Skills section.

Editing follows the same drawer, with Save in place of Create.

How a Skill is invoked at runtime​

When a Harness-mode agent receives a user request:

  1. The agent evaluates the request against the title + description of every defined skill.
  2. If the request matches a skill, the agent invokes a trigger skill tool with the matched skill's identifier.
  3. The skill executes as a sub-agent:
    • The sub-agent's base instructions are the Skill Text.
    • The sub-agent has access to the same tools as the parent agent.
    • The sub-agent processes the request and returns its result.
  4. The parent agent surfaces the result to the user.

If no skill matches, the agent handles the request directly using the Agent Profile and its own reasoning — same as ReAct mode.

Billing​

Each skill invocation incurs a fixed base credit cost per invocation (see your plan's credit configuration for the exact amount), on top of any LLM and tool usage credits consumed during the sub-agent's execution. The base cost is the same regardless of which skill ran or how briefly it ran.

In Run history and billing records, skill-triggered sub-agent runs appear with the run type label "Skill" — distinct from "Agent run". This makes skill-driven spend easy to attribute and analyse separately.

Observability​

Skill invocations are fully traceable:

  • The parent agent's trace shows the trigger skill tool call as a visible step.
  • Each skill run appears as a linked child trace under the parent agent's trace, with the skill title surfaced in the trace detail.

See Observability → Run history for how parent / child traces are presented.

When to use Harness mode​

  • Multi-purpose agents — when one conversational surface needs to handle several distinct task types (e.g., a support agent that handles password resets, billing questions, and product info — each is a Skill).
  • Reusable workflows — when the same task logic should be cleanly separated from the agent's identity, so it's easier to maintain.
  • Cleaner observability — when you want skill-level analytics (which skill is invoked most, which skill costs most, which has the highest latency).

For simpler agents where one base instruction handles everything, ReAct or Planning modes are simpler and cheaper (no per-invocation skill base cost).

Versioning and export​

Skills are part of the agent's configuration:

  • Versioning — skill definitions are captured in every saved version of the agent. See Versioning.
  • Import / Export — exporting an agent's config includes its skills. See Import / Export.
  • Archival — archiving an agent includes its skills.