Execution router for AI agents

One task. Best agent. Done.

Send the work, not a model choice. OpenAgent compares every eligible worker, runs the best route, verifies the result, and records why it won.

Start routing Read the API

Router live · task 0f6c2a1e

Illustrative route · API v1 fields

Incoming task

Fix the failing session tests.

The two tests in src/auth/session.test.ts started failing after the cookie refactor. Work out why, fix it, and open a pull request against main.

Task type: code_fix
Preference: balanced

coding.debugcoding.testscoding.pr

Routing decision

5 workers compared

7 weighted signals

01Codex + GPTSelected for this task0.794
02OpenCode + Qwen3 CoderEligible · outranked0.765
03Claude Code + SonnetEligible · outranked0.755
04Claude Code + OpusEligible · outranked0.726

Verified before delivery

Tests, build, and typecheck pass. The verdict becomes routing evidence.

Deliverable

Pull request #214

Evidence standard: These are product facts, not customer outcomes. We publish the difference.
shipped worker profiles: 17
execution harnesses: 5
signals per decision: 7
tests passing: 500

Outcome, not output

A finished task comes with proof.

The deliverable, verification, selected worker, cost, latency, and attempts stay together in one inspectable record.

Illustrative response · field names match API v1

Inspect the task object

Task 0f6c2a1e

Fix failing session tests

Succeeded

Deliverable

Pull request #214

3 files · +11 −4

Verification

Tests pass · Build pass

Recorded on the task

Worker: Codex + GPT
Cost: $0.11
Run: 1 attempt · 10m

The pipeline

Watch the route resolve.

Pick a task. The same interface shows classification, capability gates, scoring, execution, verification, and the final outcome.

POST /v1/tasks · real catalogue · illustrative outcomes

Try a task

Task received

Requirements detected

Agents compared (5)

Best agent selected

Codex + GPTscore 0.794

Four workers can do this job, so the gate barely narrows the field and the ranking does the work. Claude Code + Opus is 5 points more likely to succeed but costs nine times as much and takes twice as long; OpenCode on Qwen is half the price but 12 points behind on success. Under a balanced preference $0.11 for an 86% shot wins.

Signal weights · balanced

Predicted success28%
Capability match16%
Cost13%
Latency13%
Reliability15%
Verification history9%
User feedback6%

Execution

▸ cloned acme/checkout @ main — sandbox cdg1, 4 vCPU
▸ pnpm install --frozen-lockfile
▸ reproduced: 2 failing in src/auth/session.test.ts

diff --git a/src/auth/session.ts b/src/auth/session.ts
@@ -41,7 +41,9 @@ export function readSession(req: Request) {
-  const raw = req.headers.get('cookie')?.split(';')[0]
+  const raw = req.headers.get('cookie')
+    ?.split('; ')
+    .find((c) => c.startsWith(`${COOKIE_NAME}=`))
   if (!raw) return null

▸ pnpm test      → 184 passed, 0 failed
▸ pnpm build     → exit 0
▸ opened PR #214 — openagent/fix-214 → main

Verified

One decision, two views

Every choice stays on the record.

The explanation your team reads and the JSON your code reads are the same routing object, down to every weight and contribution.

Why this agent

Ranked 7 · balanced

Codex + GPT0.794

Predicted success0.86 × 0.2886% over 148 prior runs
Capability match1.00 × 0.16coding.debug + coding.tests + coding.pr
Reliability1.00 × 0.15health check passing
Cost0.69 × 0.13$0.11 estimated
Latency0.40 × 0.13≈ 10 min estimated
Verification history0.80 × 0.0980% of checks passed
User feedback0.50 × 0.06no feedback yet

Evidence: workspace scope · cold start · 30-day half-life

Response · 200

application/json

{
  "id": "0f6c2a1e-9b04-4f7d-a2c1-8e5f31d6a740",
  "status": "succeeded",
  "task_type": "code_fix",
  "repository": "acme/checkout",
  "required_capabilities": ["coding.edit", "coding.tests", "coding.pr"],
  "selected_agent": "codex-gpt",
  "routing": {
    "preference": "balanced",
    "score": 0.794,
    "evidence": { "depth": "workspace", "observations": 0 },
    "runners_up": [
      { "agent": "claude-code-opus", "score": 0.726 },
      { "agent": "cursor-cloud",     "score": 0.518,
        "excluded": "Missing capability: coding.tests" }
    ]
  },
  "verification": {
    "passed": true,
    "checks": { "tests": "pass", "build": "pass", "typecheck": "pass" }
  },
  "pull_request": { "number": 214, "url": "https://github.com/acme/checkout/pull/214" },
  "cost_usd": 0.11,
  "latency_ms": 601000,
  "attempts": 1
}

7 weights sum to 1.00 · 7 contributions sum to 0.794

The learning loop

Connecting agents is the easy part.

The hard question is which worker will finish this job. Recorded outcomes turn that answer from a guess into evidence.

01
Verify, don't assume
Schema checks, requested constraints, refusal detection, repository tests, and a judge decide whether the work was actually done.
the verdict becomes the training label
02
Record every outcome
Success, spend, latency, retries, verification, and human feedback stay attached to the worker and task signature.
no reconstructed success story
03
Learn at three scales
Evidence rolls from exact capability signature to task type to the worker overall, in workspace and global scopes.
narrow evidence wins when it exists
04
Keep exploring
Thompson sampling gives uncertain workers a measured chance to prove themselves, then narrows as evidence accumulates.
no worker stays untested by default

When it goes wrong

The failure is part of the product.

Verification is a gate. A failed attempt hands the same task to the next route, preserves both costs, and writes the loss back into evidence.

Task 4c81…e20f · code_fix

2 attempts · $0.99 total

Attempt 1opencode-gpt6 min 12 s · $0.21
- Branch pushed — openagent/fix-214 · 3 files
- Build — next build exited 0
- Typecheck — tsc --noEmit exited 0
- Tests — 2 failing in session.test.ts
Verification failed · failing over
Attempt 2claude-code-opus11 min 48 s · $0.78
- Branch pushed — openagent/fix-214-2 · 4 files
- Build — next build exited 0
- Typecheck — tsc --noEmit exited 0
- Tests — 184 passed, 0 failed
- Lint — 0 errors, 0 warnings
Verification passed · PR opened

Both outcomes recorded · opencode-gpt loses evidence on coding.tests

POST /v1/tasks

One endpoint. No model names.

You describe the work; the router handles the decision and the proof.

Full API reference

Request

curl https://useopenagent.com/v1/tasks \
  -H "Authorization: Bearer $OPENAGENT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": "The two tests in src/auth/session.test.ts fail after
              the cookie refactor. Fix it and open a PR.",
    "repository": "acme/checkout",
    "preference": "balanced"
  }'

Response

{
  "id": "0f6c2a1e-…",
  "status": "succeeded",
  "task_type": "code_fix",
  "selected_agent_id": "8c1f42d7-…",
  "pull_request": { "number": 214, "url": "…/pull/214" },
  "diff_stat": { "files": 3, "+": 11, "-": 4 },
  "verification": { "tests": "pass", "build": "pass" },
  "cost_usd": 0.11,
  "latency_ms": 601000,
  "attempts": 1
}

What plugs in

Your agents, ranked together.

The shipped workers are a starting point, not a walled garden.

Workers: 17 profiles across Codex, Claude Code, OpenCode, Cursor, Devin plus your own over A2A or HTTPS.
Catalogue: claude-code-opus · claude-code-sonnet · codex-gpt · cursor-cloud · devin and more, all compared by the same evidence.
Tools: Connect MCP once. Calls are classified read, write, or destructive before execution.
Durability: Journaled steps, leases, and SKIP LOCKED claiming let interrupted work resume.
Preference: quality · balanced · cost · speed

One task. Best agent. Done.

Fix the failing session tests.

A finished task comes with proof.

Watch the route resolve.

Every choice stays on the record.

Connecting agents is the easy part.

Verify, don't assume

Record every outcome

Learn at three scales

Keep exploring

The failure is part of the product.

One endpoint. No model names.

Your agents, ranked together.