Base Third Prize (ETHMumbai 2026)·Feb 2026·ETHMumbai 2026 / Paracausal Labs

Obscura: Accountable Agent-Commerce Architecture

An accountable agent-commerce architecture combining escrowed jobs, deterministic evaluation, encrypted deliverables, and persistent on-chain reputation.

Cover image for Obscura: Accountable Agent-Commerce Architecture

Overview & Research Motivation

Obscura explores an accountable gig economy for autonomous on-chain agents. Giving an AI agent economic agency should not mean granting it unconstrained authority or blind trust. Built at ETHMumbai 2026 where it was awarded 3rd Prize in the Base track, Obscura formalizes an end-to-end verifiable agent workflow.

Users specify tasks and escrow funds via ERC-8183 smart contracts on Base. Specialized agents (Scout, Analyst, Ghost) perform programmatic research and tool execution, committing encrypted deliverables to Fileverse using AES-256-GCM. A deterministic evaluator (Sentinel) verifies deliverable compliance and policy constraints before triggering escrow settlement or automated refunds, updating persistent on-chain reputation under ERC-8004.

The Core Systems Problem

  • Autonomous AI agents operating in DeFi or data research lack accountability; paying agents upfront risks non-delivery, while unconstrained execution can violate user risk policies.
  • Delivering research, trading strategies, or risk audits on-chain threatens privacy, yet completely off-chain deliveries destroy multi-party auditability and verifiable reputation building.
Formulated Question: "How can autonomous agents perform economically consequential tasks while remaining strictly bound to user-defined policies, verifiable evaluation gates, and private deliverables?"

Why I Worked on It

"I built Obscura around a simple constraint: giving an agent economic agency should not mean giving it unconstrained authority. Jobs are escrowed, agent output is evaluated against explicit policy, private deliverables remain encrypted, and successful work contributes to persistent reputation."

My Specific Technical Contributions

  • Architected the complete ERC-8183 escrow state machine and ERC-8004 reputation update logic on Base.
  • Engineered the deterministic Sentinel evaluation pipeline separating probabilistic LLM generation from financial settlement authorization.
  • Designed the per-job AES-256-GCM encryption workflow pairing symmetric key encapsulation with Fileverse decentralized storage.
  • Configured ENS text record policy schemas encoding user risk boundaries, spending limits, and emergency kill switches.

System Architecture & Verification Pipeline

ERC-8183 Escrow Lifecycle

User posts job specifications with USDC locked in escrow contracts; funds cannot move without verified completion or formal cancellation.

Deterministic Sentinel Evaluator

A hardcoded verification layer checking schema invariants, execution hashes, and ENS policy constraints without relying on probabilistic LLM judge prompts.

AES-256-GCM Encrypted Deliverables

Worker agents encrypt output payloads before storing them on Fileverse/IPFS; only the authorized employer can decrypt the full report.

ERC-8004 On-Chain Reputation

Successful job settlement updates verifiable on-chain performance records, establishing persistent agent credentials over time.

Design Decisions & Engineering Trade-Offs

Deterministic Sentinel evaluation over LLM-as-judge
Rationale: Financial settlement and fund release cannot depend on non-deterministic, prompt-injectable LLM heuristics.
Trade-Off: Requires strictly structured output schemas from worker agents.
Per-job AES-256-GCM encryption on Fileverse
Rationale: Reduces disclosure of confidential research or trading alpha while anchoring proof of delivery on-chain.
Trade-Off: Requires secure session key exchange between employer and worker agent instances.
ENS text records for user policy constraints
Rationale: Provides decentralized, self-sovereign policy management without centralized API databases.
Trade-Off: Policy updates require lightweight blockchain transaction signatures.

Evaluation & Benchmark Results

3rd Prize
Base AI × Onchain
Awarded at ETHMumbai 2026
4 Roles
Specialized Agents
Scout, Analyst, Ghost, and deterministic Sentinel evaluator

Technical Stack & Tools

SolidityFoundryBaseTypeScriptNext.jsviem / wagmiPrivyFileverseENS

Known Limitations

  • x402 protocol-version differences required custom response adapter middleware during high-concurrency test runs.
  • Sequential on-chain settlements required explicit nonce management and retry queues under network load.

What I Would Test Next

  • Formally verifying ERC-8183 escrow state transitions using Certora formal verification tools.
  • Integrating zero-knowledge computational integrity proofs for off-chain agent execution validation.

Connected Systems & Inquiries