Verifiable exception resolution for autonomous agents.
When an agent gets stuck, Resolve can buy a missing capability, verify the result, escalate when necessary, or stop safely.
Recorded UI replays and captured CLI output with edited timing. Real sandbox payments, deterministic provider payloads, and a validation-only benchmark. AI-generated narration.
Direct MP4: resolve-final-submission.mp4
An agent that cannot safely finish a task can guess, stop, or buy a more expensive capability. Resolve makes the required outcome explicit, buys a useful resolution attempt within budget, and checks the result before continuing.
AgentException → ResolutionContract → resolver selection → x402 payment
↓
agent resumes ← ResolutionReceipt ← independent SuccessContract validation
No valid receipt = no automatic resume. An answer alone is insufficient: a ResolutionReceipt requires both settled payment and independent PASS validation. Partial evidence can justify a cheaper first attempt; it cannot declare the whole contract satisfied. Failed validation triggers escalation when another eligible resolver fits the remaining budget and deadline.
| Flow | Observed behavior |
|---|---|
| Policy conflict | $850 offer → POLICY_CONFLICT → $0.001 verifier → PASS → receipt → same agent resumes at $900 without revealing a home address. |
| Capability escalation | $0.001 partial evidence → VALIDATION_FAILED → ESCALATING → $0.0015 richer evidence → PASS → receipt → resume. Total: $0.0025. |
| Economic abstention | Budget below the cheapest eligible resolver → ABSTAINED → $0 spent, no receipt, no resume. |
Payments above are sandbox settlements. The policy resolver is deterministic; the evidence providers return explicitly labeled deterministic fixtures, not live OCR/Vision responses.
The web demo follows backend SSE events. Replay observed run replays its event tape without another payment or model call. Reset demo clears presentation state, not payment history.
All three strategies evaluate the same 44 deterministic fixtures. “Verified” here means passing the deterministic validator, not completion of a paid agent task.
| Metric | CHEAPEST_ONLY | PREMIUM_ONLY | RESOLVE |
|---|---|---|---|
| Validator success | 45.5% (20/44) | 68.2% (30/44) | 68.2% (30/44) |
| Cost / verified* | $0.0017 | $0.0025 | $0.0018 |
| Total spend* | $0.0340 | $0.0760 | $0.0550 |
| Validation failures | 14 | 4 | 18 |
| Escalations | 0 | 0 | 14 |
| Abstentions | 24 | 14 | 14 |
* Hypothetical resolver quotes. The benchmark executes no payments, issues no receipts, and resumes no agents. Cost per verified result includes failed attempts. Validation failures count attempts; an escalated case can fail once and then pass. Zero unverified resumes in this harness is not evidence of live resume safety.
Benchmark output · Recorded live readiness checks · Sandbox payment proof
flowchart TD
A[OpenAI agent: paused run] --> B[AgentException]
subgraph Resolve Runtime
B --> C[ResolutionContract]
C --> D[Planner / BudgetGuard]
D --> E[Selected resolver / x402 sandbox payment]
E --> F[Resolver output]
F --> G[Independent SuccessContract validator]
G -->|FAIL: budget and deadline permit| D
G -->|PASS + SETTLED| H[ResolutionReceipt]
D -->|No eligible attempt| I[ABSTAINED]
end
H --> J[Original agent resumes with verified result]
TypeScript, Node.js, Express, Next.js/React, Zod, OpenAI Agents JS, x402/Solana, SSE and Vitest. Settlement is checked against the payment response and RPC transaction evidence. Agent messages go to a local transcript, not to a buyer.
Use Node.js 22.12+ to satisfy dependency engine requirements.
git clone https://github.com/ikrrispatel/resolve.git
cd resolve
npm install
npm run demo:policy
npm run demo:escalate
npm run demo:abstain
npm run benchmarkThe three demo commands default to core-only payment/validation runs; they do not claim agent continuation. Sandbox calls need network access and use disposable in-memory signers. No Pay account or mainnet wallet is required. The benchmark needs neither credentials nor payment access.
For the web interface and live agent, copy .env.example to .env and set OPENAI_API_KEY locally. Never commit it. OPENAI_MODEL is optional and defaults to gpt-4.1-mini; model calls use your API account.
npm run devOpen /demo or /benchmark on the local web server at port 3000; the API uses port 4310. Only one web demo run is active at a time. Add -- --agent to a demo command for the real OpenAI pause/resume path, or -- --verbose for full event details.
npm test
npm run typecheck
npm run buildHosted frontends require RESOLVE_API_ORIGIN at build time and a separately running API configured with RESOLVE_FRONTEND_ORIGIN. Keep OPENAI_API_KEY on the API server, never in a client environment variable. No public live deployment is claimed here.
apps/web/ Next.js execution demo and benchmark
apps/api/ Express runtime and SSE
packages/core/ Contracts, planner, BudgetGuard, validation and events
packages/payments/ x402/SVM settlement verification
packages/agent-runtime/ OpenAI interruption and receipt-gated continuation
packages/resolvers/ Policy verifier, evidence fixtures and catalog adapters
packages/benchmark/ Shared fixture evaluation
scripts/ CLI demos, benchmark and verification
docs/ Design documents and presentation assets
Design sources, in priority order: AI_RULES, PRD, Architecture, PLAN, REFERENCE_REPOS, HYPER_PROMPT. These describe the original plan; the implementation and limitations above describe the shipped demo.
- Real sandbox x402 payment settlement.
- Independent validation before resume.
- Receipt-gated continuation of the same OpenAI agent run.
- Automatic escalation after partial evidence fails validation.
- Budget-based abstention with no payment or resume.
- Mainnet catalog settlement or production-scale reliability.
- Authenticity detection from deterministic evidence fixtures.
- Real-world provider economics from the validation-only benchmark.
Live runs depend on hosted sandbox and OpenAI availability. Run state is process-local and disappears on restart; SQLite is deferred. Receipts are internal records, not externally authenticated credentials for arbitrary clients. Production access controls and a deployed marketplace are outside this prototype.
MIT.