T3N Trusted Approval Agent

Problem

Independent professionals lose time checking fragmented task marketplaces, verifying whether rewards and eligibility are real, and deciding whether a task is safe to claim. Blind automation is dangerous: it can accept unclear legal terms, expose credentials, trigger payments, or claim work that cannot be completed.

Solution

T3N combines a task hunter with an approval gateway and tamper-evident audit log. Each opportunity is normalized, evaluated against explicit evidence, and either rejected, escalated for review, or claimed under the user's standing approval. The agent fails closed when reward, geography, AI-use permission, payout safety, or zero-cost execution cannot be verified.

Why This Matters

The system gives solo professionals more reach without surrendering control. It automates repetitive discovery and verification while preserving human authority over consequential or ambiguous actions.

How We Used AI

The project uses the Strands Agents TypeScript SDK with two typed tools: evaluate_paid_task and start_eligible_task. The model can reason about a listing and invoke the workflow, but it cannot bypass deterministic policy. Tool calls prohibit fabricated claims, CAPTCHA solving, credential collection, KYC handling, payments, or starting a task with missing evidence.

How We Used Codex

Codex inspected the existing safety gateway, integrated the Strands SDK, designed the typed tool boundary, added the standing-approval rule, expanded task-evidence validation, wrote tests, ran the end-to-end demo, and published the verified implementation to GitHub.

Key Features

Evidence-based task screening for AI permission, confirmed reward, Bulgarian eligibility, zero upfront cost, legal clarity, and safe payout. Deterministic fail-closed policy beneath the AI layer. Standing approval for task.claim only after all hard gates pass. Explicit review path for ambiguity or elevated risk. Hash-chained audit records with integrity verification. Safe demo covering both an accepted SEPA task and a blocked KYC/crypto task.

Architecture

Marketplace adapters feed normalized opportunities into the Strands agent. Its typed evaluation tool calls the deterministic task hunter. Eligible work proceeds through the trusted approval gateway to a claim action; anything ambiguous stops at review. Every decision is appended to the audit chain.

Architecture diagram: docs/t3n-architecture.png in the public repository.

Testing Instructions npm install npm run check npm run demo:strands

Expected result: six tests pass. The verified zero-cost SEPA example reaches CLAIMED; the unconfirmed KYC/crypto example reaches REVIEW_REQUIRED; the audit chain reports valid.

Optional live-model run (requires configured AWS credentials and model access): RUN_STRANDS_MODEL=1 npm run demo:strands

Public Demo Link https://t3n-trusted-approval-agent.kostovdobromir.chatgpt.site