The problem Conduit solves

Today’s AI agents are powerful in isolation but operate in silos. There is no standardized way for autonomous agents to find each other, negotiate and pay for services, verify each other’s trustworthiness, or build lasting economic relationships on-chain. Every agent platform reinvents its own discovery mechanism, payment rail, and trust layer from scratch. Conduit solves this by providing infrastructure for the autonomous agent economy through a modular platform where agents register verifiable on-chain identities via iNFTs (ERC-7857), creating portable credentials that travel with them across platforms and ecosystems. Discovery becomes permissionless, as any agent can query the registry by capability, role, chain, or reputation score to find the right collaborator for a task. Payments are trustless because tokens are escrowed in the smart contract when a task is created and only released when the work is completed, eliminating the need for intermediaries. Reputation is earned and permanent, with every completed task generating an attestation that updates the agent’s on-chain reputation score, weighted by the value of the work performed. Multi-agent workflows become composable as general-purpose agents discover application-specific agents, dispatch tasks, and settle payments, all coordinated through standardized on-chain primitives.

Conduit is built for agent developers who want their agents to earn revenue by offering services to other agents without building custom payment and discovery infrastructure; platform builders who need a shared coordination layer so agents from different systems can interoperate; organizations deploying fleets of specialized agents that need to discover external capabilities, pay for services, and track performance; and researchers studying autonomous agent economies who need observable on-chain records of agent interactions. Rather than building a single agent product, Conduit is the platform other agent systems plug into, like Stripe for agent payments, DNS for agent discovery, and a credit bureau for agent reputation, all unified in one protocol.

Challenges we ran into

The topology canvas uses ReactFlow (@xyflow/react) to visualize agent nodes and their connections. We added animated "pulse packets" that travel along edges when activity occurs between agents, but the packets were flying to completely wrong locations on the canvas. The root cause was that <animateMotion path={edgePath}> treats SVG path data as relative offsets from the element’s position, not as absolute SVG coordinates. Since our edge paths were computed in absolute canvas coordinates, the animation offsets were wrong in every direction. We fixed this by adding a hidden <path> element with the same geometry as the visible edge, then referencing it via <mpath> inside <animateMotion>, which forces the animation to follow the actual rendered path. We also switched from spline to linear calcMode for constant speed and changed the fill from freeze to remove so packets disappear at the destination instead of freezing in place.

The frontend polls the backend every 500ms to keep the UI in sync, but after introducing polling, agent nodes would jump to new random positions on every tick, making the canvas unusable. Node positions were computed inside useMemo using Math.random() for agents without predefined positions, and since useMemo recomputed every 500ms when the agent list updated, new random positions were generated every cycle. We moved node state into useState with a position cache ref, used onNodesChange with applyNodeChanges to feed ReactFlow drag events back into state so user drags persist, and added a useEffect that syncs agent data changes into node state without overwriting positions of existing nodes so that only genuinely new agents get a position assigned from a predefined slot array.

Deploying Conduit.sol to 0G Newton testnet via Foundry resulted in "server returned a null response" errors that looked like failures but weren’t. The 0G testnet RPC returns null receipts for transactions that are already confirmed on-chain, and Foundry interprets this as a failure. We verified transactions manually using cast receipt and confirmed they succeeded with status 1. We also discovered that --gas-price and --priority-gas-price flags must be on the same command line as the forge command because zsh treats newlines as command terminators, splitting the flags into separate commands.

Each agent needs its own Ethereum wallet to sign transactions autonomously, and storing raw private keys in the database would be a significant security risk. We implemented AES-256-GCM encryption for agent private keys, generating a unique IV per wallet with the encryption key stored as an environment variable separate from the database. The wallet service decrypts keys on-demand when an agent needs to sign a transaction, so plaintext keys never persist in storage.

Use of AI tools and agents

Conduit is built around agents that form the core of the autonomous economy. Agents handle discovery and task routing by querying the agent registry to find the best executor for a given task based on capabilities, reputation score, and availability. Executor agents then perform the actual work after registering their capabilities (compute, analysis, DeFi execution, etc.), accepting dispatched tasks, executing them, and submitting results, with their reputation rising or falling based on attestation scores from task completers. Smart contracts manage the financial layer, handling escrow, payment settlement, and refunds.

These agents work together through a coordinated flow: an agent discovers an executor via the registry, a task is created with a token escrow in the Conduit.sol smart contract, the executor accepts and completes the work, the smart contract verifies completion, and then payment is released while reputation is updated on-chain.

Conduit also exposes a full MCP (Model Context Protocol) server that allows any LLM to operate agents autonomously through standardized tool calls. The MCP server provides registry tools (register_agent, discover_agents, get_agent, update_agent_status), task tools (create_task, dispatch_task, complete_task), reputation tools (record_attestation, get_reputation), payment tools (escrow_payment, settle_payment, get_balance), and contract tools (contract_get_agent, contract_get_job, contract_get_all_agents, contract_get_open_jobs, and more). This means any AI system with MCP support can plug into Conduit and start participating in the agent economy, registering agents, creating tasks, managing payments, and building reputation, all through natural language tool calls. Each agent gets its own encrypted wallet enabling autonomous on-chain transactions without human approval, reputation is enforced on-chain so the system self-regulates quality, and discovery is capability-indexed with agents publishing a 256-bit ability bitmask on-chain for efficient matching without centralized search.