AgentClash Agent Harness Setup

Purpose

Configure and operate Agent Harnesses — workspace-scoped autonomous coding tasks with E2B runners, evaluation config, suite rankings, and failure curation. Agent Harnesses are not challenge packs.

Use When

• A user wants long-running coding-agent checks against a repository with validators or LLM judges.
• You need to create harnesses for Codex, Claude, Hermes, or OpenClaw runners on E2B.
• The workflow involves suite task banks, multi-harness suite runs, or promoting failed executions into private tasks.

Do Not Use When

• The workload is a standard challenge-pack eval — use agentclash-eval-runner.
• The user only needs deployments or runtime resources — use agent-build skills first.
• The task is prompt A/B testing without a repo harness — use agentclash-prompt-eval-playground.

Inputs Needed

• Workspace with provider secrets configured (openai_api_key_secret_name or --api-key-secret).
• Task prompt, repository URL, and harness kind.
• Optional evaluation config JSON (validators, LLM judges).
• For suite runs: suite ID, harness IDs, optional task filters.

Environment

export AGENTCLASH_API_URL="https://api.agentclash.dev"
agentclash workspace use <WORKSPACE_ID>
agentclash secret list --json

Procedure

1. List or inspect existing harnesses.
2. Create a harness with --name, --task, --auth-mode, and API key secret.
3. Run a harness execution; use --follow to poll to terminal status.
4. Inspect executions, failure summaries, and failure reviews.
5. Optionally create suites, run across harnesses, read rankings, promote executions to private tasks.

Commands

Harness CRUD and runs

agentclash agent-harness list
agentclash agent-harness get <harness-id>
agentclash agent-harness create \
  --name "Refund fix" \
  --task "Fix the refund bug in services/refund.go" \
  --harness-kind codex_e2b \
  --auth-mode api_key_secret \
  --api-key-secret OPENAI_API_KEY \
  --repository-url https://github.com/org/repo \
  --base-branch main

agentclash agent-harness run <harness-id> --follow
agentclash agent-harness executions <harness-id>

--harness-kind values: codex_e2b (default), claude_e2b, hermes_e2b, openclaw_e2b.

Create from JSON:

agentclash agent-harness create --from-file harness.json

Optional flags: --codex-template, --codex-model, --execution-config, --evaluation-config, --evaluation-config-file.

Executions

agentclash agent-harness execution get <execution-id>
agentclash agent-harness execution cancel <execution-id>
agentclash agent-harness execution retry <execution-id> --idempotency-key cli-retry-1

Suites and rankings

agentclash agent-harness suite list
agentclash agent-harness suite create --name "Private bank" --task-json '{"title":"Task 1","public_prompt":"..."}'
agentclash agent-harness suite tasks <suite-id>
agentclash agent-harness suite run <suite-id> --harness <harness-id-1> --harness <harness-id-2>
agentclash agent-harness suite rankings <suite-id> --k 3

Failures and promotion

agentclash agent-harness failures summary
agentclash agent-harness execution failure-review get <execution-id>
agentclash agent-harness execution failure-review update <execution-id> --human-class timeout --human-summary "Sandbox timed out"
agentclash agent-harness execution promote-task <execution-id> --suite <suite-id> --title "Promoted failure case"

Alias: agentclash harness = agentclash agent-harness.

Expected Output

• Create returns harness ID, kind, auth mode, template.
• Run returns execution ID; --follow polls until completed, failed, or cancelled.
• Suite rankings table shows success@1, pass@k, cost, latency per harness.

Failure Modes

• Missing --api-key-secret on create → pass workspace secret name containing provider key.
• Missing required suite flags → --harness required on suite run; --task-json or --from-file on suite create.
• Non-terminal execution on retry → only terminal executions can retry.
• Wrong harness kind for template → defaults: codex, agentclash-claude-fullstack, agentclash-hermes-fullstack, agentclash-openclaw-fullstack.

Safety Notes

• Harness runs execute code in E2B sandboxes with repository access — confirm repo and secrets before running.
• Promoted tasks may contain sensitive failure excerpts — sanitize public_prompt.
• Do not paste API keys or secret values into chat.

Report Back Format

Harness: <id> (<name>, <kind>)
Execution: <id> — <status>
Suite: <id or n/a>
Ranking summary: <top harness or n/a>
Failure review: <effective_class or n/a>
Next commands: <1-3>

Related Skills

• agentclash-hub
• agentclash-cli-setup
• agentclash-runtime-resources-setup
• agentclash-eval-runner
• agentclash-scorecard-reader
• agentclash-regression-flywheel

Related Docs

• /docs-md/guides/ci-cd-workload-recipes
• /docs-md/reference/cli