What is the problem your project solves?

Edge AI agents crash or stall when they generate code they can't actually run — not enough RAM, missing dependencies, no pre-configured cloud sandbox. Today the only recovery paths require a human: wait, provision cloud infrastructure in advance, or give up on the task. There's no way for an agent to autonomously recover from a resource constraint at the moment it happens. This matters because real autonomy requires agents to handle failure without a human in the loop — and right now, "ran out of local compute" is a hard stop, not a recoverable state. The impact: any resource-constrained device — edge hardware, a laptop, a Raspberry Pi — could instead borrow spare compute from a nearby peer, on demand, with no pre-existing trust relationship or cloud account.

How you are solving it?

Overflow is a peer-to-peer compute delegation protocol. Built on top of Edgent (github.com/R-Abinav/edgent, prior work from ETHGlobal Open Agents 2026, disclosed in full), which already solved encrypted P2P peer discovery over the Gensyn AXL mesh, staking, and payment settlement — but only for forwarding a single inference prompt to a peer.

During this hackathon, we extended that into real sandboxed code execution. A requester agent generates code locally; if it hits a resource constraint, it delegates the actual code — not just a prompt — to a peer over the mesh. The peer runs it in a stdlib-only Python sandbox with no pre-installed dependencies, so any missing-library failure is genuine, not staged. If it fails, the failure is fed back and self-corrected in a real retry loop. The entire core pipeline (sandbox.ts, the message schema) is objective-agnostic — it has no idea what task it's running.

We then built an SRE reference client on top: it reads a real production incident log, delegates diagnostic work when resource-constrained, gets back a root-cause diagnosis grounded in the actual log data, generates a patch, and opens a real GitHub pull request — automatically, with zero human intervention after the log is dropped in. During the build we also caught and fixed several real correctness bugs (a hardcoded local-execution stub, a timestamp-parsing failure, an output-fidelity bug where cited numbers didn't match computed data, and a validation gap that let a failed task cascade into a fabricated PR) — all disclosed and fixed with verified before/after evidence in the commit history.

How Did You Use Claude?

Claude is the core intelligence layer on the provider side. Every delegated task is handed to a Claude agent (claude-sonnet-5 via the Anthropic Messages API in src/core/llm.ts) that writes a single self-contained Python script and executes it. When the script fails on a real error — e.g. reaching for a package that isn't installed in the sandbox — Claude reads the actual stderr and rewrites the script until it passes or the attempt budget runs out. That loop originally ran on Gemini during early testing and was swapped to Claude mid-hackathon after re-verifying the full test matrix (commit 299f8ee).

Claude also powers the RCA (root-cause-analysis) pipeline: given a failure log, it diagnoses the root cause, generates a patch, and opens a real GitHub pull request from that patch — with an output-validation guard added during the hackathon so a malformed or errored diagnosis step halts the pipeline instead of cascading into a bad patch.

What is the deployed URL for this project?

https://github.com/R-Abinav/overflow