Summary

HarnessCard is a lightweight reporting artifact proposed by He et al. (2026) for disclosing the harness configuration of language agent systems. Similar to how Model Cards document model properties, HarnessCard documents the apparatus that makes an agent claim interpretable, comparable, and reproducible.

Key Points

Purpose

  • Make agent claims comparable across different systems
  • Enable auditable harness configurations
  • Support reproducible agent evaluations
  • Reveal the harness-sensitive component of reported gains

Required Fields

FieldWhat to Disclose
Base model(s)Model name, version, decoding settings, finetuning or adapters
Control artifactsSystem instructions, AGENTS.md, repo maps, architecture rules, schemas, tests, linters, done-when criteria
Runtime policyMemory type, compaction/summarization policy, checkpointing, retry/rollback policy, budget limits
Action substrateTools, APIs, browser/GUI access, code execution, interface schemas, MCP usage
Execution topologySingle vs multi-agent structure, planner/verifier roles, reviewer loops, routing logic
Feedback stackTests, graders, reflection prompts, hidden checks, human interventions, repair loops
Governance layerPermissions, sandboxing, escalation rules, policy checks, provenance logging, audit support
ObservabilityStored traces, replay support, latency and cost logging, failure categories
Evaluation protocolTask set, number of runs, success criteria, variance treatment, held-out checks, budget limits
FieldWhat to Disclose
Release artifactsPrompts or programs, tool specs, traces, configs, environment setup, reproducibility notes
Known limitationsUnresolved failure modes, portability caveats, safety concerns, red-team findings

Example: Repository Coding Agent

FieldIllustrative Disclosure
Base model(s)Frontier coding model configured through repo or user profiles; effort tuned for long tasks
Control artifactsRoot-level AGENTS.md; repository map; build/test/lint commands; architecture rules; done-when criteria
Runtime policyRepository treated as system of record; thread history; progress file; compaction near context limits; bounded retries
Action substrateFile edits, shell commands, test runs, diff generation, PR review, optional MCP tools
Execution topologyPlan → edit → run tools → observe → repair → update status → repeat; optional reviewer loop
Feedback stackFailing tests, custom linter messages, self-review, grader checks, occasional human review
Governance layerSandbox mode, approval policy for privileged actions, least-privilege connectors, audit trail
ObservabilityPersisted thread events, replay support, latency and cost logs, categorized failures
Evaluation protocolStandard coding benchmark; 3 runs per task; pass@1 and pass@3; variance reported

Why It Matters

The Visibility Gap

He et al. audit 63 harness-relevant works and find:

  • Academic papers often treat harness as “hidden implementation residue”
  • Public engineering notes (Anthropic, OpenAI) describe innovations not yet in papers
  • Without HarnessCard, it’s unclear whether gains come from model or harness

Comparison with Model Cards

AspectModel CardHarnessCard
FocusModel propertiesAgent system apparatus
DisclosureTraining data, biases, capabilitiesControl, agency, runtime configuration
GoalTransparency about modelReproducibility of agent claims
Analogy”What is this model?""How does this agent work?”

Relationship to Other Concepts

  • CAR Framework: HarnessCard documents each layer of ⟨C, A, R⟩
  • Harness Engineering: HarnessCard is the reporting artifact of harness engineering practice
  • Agent Evaluation: HarnessCard makes evaluation claims interpretable and comparable

Open Questions

  • What fields should be truly required vs recommended?
  • How to handle proprietary harnesses that can’t fully disclose?
  • Should HarnessCard become a community standard or remain a proposal?
  • How does HarnessCard interact with evaluation frameworks (HAL, ATBench, VeRO)?

Evidence Timeline

  • 2026-04-23: Created from He et al. “Harness Engineering for Language Agents” (Alibaba-NTU, Preprints.org)
  • 2026-04-23: HarnessCard proposed as lightweight reporting artifact with 9 required + 2 recommended fields
  • 2026-04-23: Purpose: make agent claims comparable, auditable, reproducible

相关页面

sandbox