The role

Payle's authorization engine is live: agents ask for permission, a deterministic policy engine decides, every transaction lands in a hash-chained ledger that anyone can verify. The money path is audited, money-tested under concurrency, and finished.

What's missing is the layer agents run on. You would own it end to end: the agent runtime that searches real retailers, captures evidence, compares offers, attempts purchases through the authorization gate, and verifies outcomes. Your work is the reason the gate exists: without the runtime, the engine is a vault with no hands.

What exists

• The Go authorization core: policy DSL, idempotent under concurrent agent retries (proven: 100 parallel identical requests, exactly one decision), hash-chained append-only ledger, kill switch, reconciliation. Externally audited; findings fixed with regression tests in CI.
• A live demo: our agent buys real domains through the gate at mattiaciuni.pages.dev/agent: real search, real policy decline, real receipts.
• The OpenAPI contract between engine and runtime: your layer consumes it, and helps us evolve it.

What you'd build

• The full agent runtime: planning loops with frontier LLMs (tool calling, structured outputs), search connectors across retailers (Apple, Amazon, Back Market for the launch campaign; broader catalog over time), a browser pipeline for evidence capture (screenshots, extraction hashes, freshness checks)
• Deterministic comparison and ranking: same input, same ranking, every time: with written explanations for every recommendation. The LLM plans; it never ranks silently
• The MacBook Gate: our launch campaign where the agent tries to buy a MacBook 2,000 times under policy, declines every time, and buys one real one for the draw winner. Your runtime powers it
• Outcome verification: the charge is not the task. You build the checks that prove the domain registered, the order shipped, the subscription cancelled

You

• TypeScript/Node strong: the runtime lives in TypeScript. Go familiarity is a plus (the engine you'll talk to is Go)
• LLM orchestration in production: tool calling, structured outputs, planning loops: and you know where agents break: retries that double-charge, silent prompt injection, workflows that look fine in demo and die at scale
• Web automation real experience: Playwright, anti-bot realities, evidence capture without faking anything
• You understand money: idempotency, the difference between a timeout and a decline, why it worked in the demo is not a deployment standard
• Written communication: the team is distributed and writes everything down. You explain decisions in text, cleanly

Compensation

€2,500-3,000/month + 0.75-1% equity (4-year vesting, 1-year cliff). Contractor now, full-time at YC acceptance, relocation possible with sponsorship.

The challenge

Your first artifact: a small agent with a hard boundary. Build an agent that does useful work while keeping authorization deterministic and inspectable. Deliverable: a working repository, a short decision log, and tests that show where the boundary holds. We pay for your time.