I build AI agents, then write down exactly where they break.
An AI agent's own account of a ten-hour investigation: a photo of a bench with no location data, five confident wrong answers, a decrypted four-year message archive, and the architectural defects the whole thing exposed. Written by the agent that ran the session; the corrections are from me.
I work in the open at Zion-sewers — public infrastructure for trust in agents. Every repo answers one question: how do I know I am right?
- mcp-guard — adversarial fuzzer for MCP servers. 65 findings against Anthropic's reference servers, including silent shell-injection hits.
- sqltok — token-budgeted schema selection for Text-to-SQL. 97% gold-table recall at a 2,000-token budget on BIRD.
- agenteval-bench — pytest for agent outputs. Deterministic matchers, regression tracking, CI gates.
- protobuf-fuzz-guard — protobuf schema CVE scanning plus fuzz-harness generation in four languages.
The live Zion board. Every card, every office, zero hidden queues. This is where the work above actually happens.
I think 2027 will be around optimizing around human related metadata for example creating your work insights based on your data and autonomous scheduling acts on it. Chat-interface is decaying, and workflows styled agentic is the main meta approach, but internal business end-user implementations are difficult because of popular global trends: AI Jobless boom spoken by none other of Geoffrey Hinton, a living fossil aka Godfather of AI.
Currently handling multi-eval,centralized MCP, video generation flows, and A2A across sys computer used agents cloud based scheduling. Handling and learning about Agentic security in Code, pipeline(devops),resource/deployed bicep(terraform), Data level (e.g credit card leaked in transcription, and Agent token level - guardrails, prompt instructions,structured hierarchal tooling).
As for 2026 I’m still scared and aroused from mathematical notations and formalization. Ms suggestions/research colabs are always welcomed in dm.
Tel Aviv. More writing at daily-deep-learning. This profile is a work in progress.