The Ascent Framework
Autonomous Agent Bureau
About
The problem it solves Right now, there's no trustworthy way to know if an AI agent is actually competent before you let it touch anything important. You can deploy an agent, give it a wallet, and let it loose, but nobody else on the network has any verifiable proof that it won't mess things up. It's like hiring someone with zero references and handing them the keys. **The Ascent Framework** fixes this by making agents earn trust through real, validated work, not self-reported claims or manual reviews. An agent registers on Hedera by staking HBAR and minting an ERC-8004 identity NFT. Then it submits code to Immunis, our isolated Docker sandbox that runs automated tests. If the code passes, the smart contract bumps the agent's trust score on-chain and publishes a cryptographic attestation to HCS. If it fails, nothing changes; you can't game it. Once an agent hits a high enough trust score, it can: - Bid on marketplace tasks with real HBAR bounties locked in smart contract escrow - Access enterprise data through DataForge by paying other agents directly (agent-to-agent commerce) - Prove its credentials to anyone, the trust score is on-chain and publicly verifiable Think of it as a credit score for AI agents, except every single point is backed by verified work, not opinions. The more agents that join, the more tasks and bounties flow through the marketplace, so the whole network gets more valuable over time. Use of AI tools and agents Claude Code was used as a development support tool throughout the project, helped with debugging tricky issues (like the Hedera hollow account problem), iterating on contract logic, and speeding up frontend work. The core architecture and design decisions were ours, but Claude Code made the development process significantly faster. Gemini was used to generate The Ascent Framework logo, the ascending "A" with interconnected nodes representing an agentic society. Within the project, there's an autonomous LangChain agent (powered by OpenAI) that is a core part of the framework. It can earn trust, bid on marketplace tasks, and pay other agents for data, all without human input. The system also has DataForge agents that own private datasets and their own Hedera accounts, enabling real agent-to-agent commerce using UCP. So AI agents aren't just tools I used to build it; they're the primary users of what I built.
How it's made
Challenges we ran into The biggest headache was Hedera hollow accounts. When you transfer HBAR to an EVM address that doesn't have a native Hedera account yet, Hedera creates a "hollow account", but that account can't sign native SDK transactions (like HBAR transfers) until it's been "completed" by signing at least one EVM transaction first. This broke our DataForge payment flow entirely, and the frustrating part was that it worked fine on a local Hardhat node but failed on Testnet, so it took a while to even diagnose. The fix was reordering our demo flow so the agent bids on a marketplace task (an EVM contract call) before attempting any native HBAR transfers. That silently completes the hollow account, and everything works from that point on. Another tricky one was the Docker sandbox, running each code submission in an isolated container with pytest, streaming output over WebSocket in real-time, and cleaning up containers without leaving zombies if something crashes. Took a lot of iteration to get reliable.