A 6-month, $25,000 project building a failure-attribution harness and public benchmark for multi-agent AI systems.
Scythe is open tooling for attributing failures in multi-agent AI systems. Multi-agent AI systems are being deployed faster than our ability to understand why they fail. When several agents share tools, memory, and a task, a bad outcome can come from a bad plan, a misread tool result, corrupted shared state, or one agent misleading another, and current logs do not say which. Without attribution, teams cannot fix the real cause, cannot tell whether a failure was an accident or a coordination breakdown, and cannot build safety cases for systems that act on their own for hours at a time. As agents get longer horizons and more autonomy, these failure modes become mechanisms of large-scale harm. This is a 6-month, part-time project by Sai Nitish Reddy Kapa, based in Philadelphia, PA, USA. Funding goal: $25,000.
(1) A failure-attribution harness for multi-agent runs: instrumented traces recording each agent's plan, tool calls, tool results, and shared-state changes, classifying a failed run into planning error, tool misuse, misread observation, corrupted shared state, or coordination breakdown between agents. (2) A public benchmark of scripted multi-agent failure scenarios with ground-truth causes, drawn in part from real open-source multi-agent failures. (3) A written study on coordination-breakdown failure modes between agents. (4) A written report and open-source release of the harness and benchmark, aimed at small teams building multi-agent systems who have no safety tooling today.
The harness will classify failed multi-agent runs into the taxonomy on scripted scenarios with known causes. The benchmark will be public with ground-truth causes so others can check and reuse it. The coordination-breakdown study and report will be published, and the harness and benchmark will be released open-source.
$25,000 total: $20,000 time support (6 months, part-time) + $5,000 compute and model API costs for running benchmark scenarios.
Sai Nitish Reddy Kapa, Philadelphia, PA, USA. M.S. Data Science, University at Buffalo SUNY, May 2024. 5 years Python; 3 years NLP/LLM including fine-tuning; RAG, vector databases, retrieval pipelines, agent frameworks, LangChain, LangGraph; built Claude Code plugins and MCPs. Founder of Syrava; built the Syrava Everyday MCP server (13 tools), Scythe (Agentic Code-Review Workflow), and Vakya (Browser-Based Speech Workflow, 117 automated tests). Forward Deployed Engineer (freelance) Jan-Sep 2026; Business Data Analyst, Augmentare Inc. Applied Nov 18, 2026 as a team member for AI Safety Camp 12th edition (Jan-Apr 2027), first-choice project Agent Forensics: Attributing Failures in Multi-Agent AI Systems. Links: https://github.com/kapasainitishreddy and https://kapasainitishreddy.github.io/portfolio
The attribution taxonomy may not generalise beyond scripted scenarios; I will mitigate this with scenarios drawn in part from real open-source multi-agent failures and by stating the limits plainly in the report. I am a solo builder; I will mitigate this via the open-source release and the AI Safety Camp connection for review from experienced researchers. If the project fails, the likely outcome is a narrower harness and benchmark than planned, with the limits documented in the report.
Nothing raised for this project in the last 12 months. Note: the same project was also applied for on Nov 18, 2026 to BlueDot Impact Rapid Grants ($8,000, smaller 3-month version) and the EA Funds Transformative AI Fund ($25,000), both awaiting decision.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!