B.6 Harness composition Active

Given a task family, the agent assembles skills, plugins, tools and MCP servers into a harness that completes it. Graded by the score difference against other harnesses on the same model and by completeness of the harness card.

Evidence: none. No benchmark card lists this capability yet. No level is issued: methodology 0.1.0 sets no level thresholds.

Domain: B · Embodiment (tools)
Evidence form: program
Veto constraints: —
Depends on: B.2 Tool registration B.3 Tool invocation B.4 Tool result handling
Required by: —
Internal task families: —
External anchors: tau-bench, SkillsBench
Version: 0.1.0 · 0.1.0

Definition in detail

Scores on the same model differ markedly between harnesses, so the harness is part of what is measured. Composing the right components for a task family is a capability in its own right.

Inputs a task gives

A task family described at the level of method
A catalogue of available skills, plugins, tools and MCP servers
The harness card schema
Constraints on permissions and resources

Artifact required

A harness definition (components with versions, configuration and invocation policy) and a harness card stating the model slot, components, permissions, autonomy level and known limitations.

How it is graded

A program runs the task family with the composed harness and compares its score with other registered harnesses on the same model. The harness card is checked for completeness against its schema and for consistency with the components actually loaded.

In scope

Selecting and configuring components for a task family
Declaring permissions and autonomy accurately
Documenting the harness so that a run can be reproduced

Out of scope

The quality of any single tool - graded under B.1 to B.5
Capability of the underlying model independent of the harness

Typical failures

Expected failure patterns; not measured rates.

Loading components with overlapping functions that issue conflicting calls
A card that lists components not actually loaded, or omits ones that are
Granting broader permissions than the task family needs

Standards: std-agent-skills · std-mcp-2026-07-28 · std-harbor-task-format

Benchmarks

No benchmark card lists this capability.

Grading references

No grading reference is mapped to this capability.

Levels by model and harness

No verified results yetLevels would be derived from site-verified result records; methodology 0.1.0 sets no level thresholds, so the level shows not issued.