Deebo is a fully autonomous debugging agent that integrates into LLM workflows via the Model Context Protocol (MCP).
When a coding agent like Claude hits a bug it can’t confidently fix, it offloads the problem to Deebo.
Deebo’s “mother agent” spawns multiple scenario subprocesses—each exploring a different hypothesis in its own Git branch. These subprocesses use tools like git-mcp and desktopCommander to validate or falsify their fixes. Results are reported back asynchronously, and the mother agent converges on the best solution through structured reasoning loops.
To stress-test it, we picked a real-world $100 bounty from george hotz' repo, tinygrad: test_failure_53. The bug was buried deep in the Linearizer optimizer. Several devs had taken a stab at it, but the root cause wasn’t clear.
Here’s what Deebo did:
• Parsed the test and repo structure
• Hypothesized failure due to GROUPTOP optimization on uchar reductions
• Explored type promotion issues, stride edge cases, and scope conflicts
• Proposed a valid fix: disable GROUPTOP in unsafe contexts or add explicit type coercion
The test passed. The fix is still under review—but the important part is: Deebo figured this out on its own. It didn’t copy-paste code or regurgitate docs—it ran structured experiments, refined hypotheses, and converged on the solution path.
If you’re working on agents, devtools, or complex debugging workflows—we’d love feedback. AMA about the architecture, MCP integration, or what’s next.
1 comment
[ 3.5 ms ] story [ 10.6 ms ] threadDeebo’s “mother agent” spawns multiple scenario subprocesses—each exploring a different hypothesis in its own Git branch. These subprocesses use tools like git-mcp and desktopCommander to validate or falsify their fixes. Results are reported back asynchronously, and the mother agent converges on the best solution through structured reasoning loops.
To stress-test it, we picked a real-world $100 bounty from george hotz' repo, tinygrad: test_failure_53. The bug was buried deep in the Linearizer optimizer. Several devs had taken a stab at it, but the root cause wasn’t clear.
Here’s what Deebo did:
• Parsed the test and repo structure
• Hypothesized failure due to GROUPTOP optimization on uchar reductions
• Explored type promotion issues, stride edge cases, and scope conflicts
• Proposed a valid fix: disable GROUPTOP in unsafe contexts or add explicit type coercion
The test passed. The fix is still under review—but the important part is: Deebo figured this out on its own. It didn’t copy-paste code or regurgitate docs—it ran structured experiments, refined hypotheses, and converged on the solution path.
If you’re working on agents, devtools, or complex debugging workflows—we’d love feedback. AMA about the architecture, MCP integration, or what’s next.