Carbonforge: validation is the bottleneck for agentic discovery; its first paper hits a NeurIPS workshop
pierrelux · x · 2026-10-15
Pierre Lux of Carbonforge (@carbonforgeai) shares that with their lab in the loop, agents now continuously improve their stack end-to-end. But the real bottleneck in agentic systems is validation—confirming whether what an agent found is a real solution, true for code and math proofs alike.
They built several mechanisms into their harness to counter false discoveries (e.g., EvidenceTree, an MCP gate/ledger recording evidence and claims). He concedes it's not infallible: an agent could bypass the MCP or even tamper with the ledger, though that would surface in an audit. The company's first paper will be presented at the NeurIPS e-values workshop in Paris this December—e-values support anytime-valid inference and continuous monitoring, well-suited to sequentially evaluated ML systems.
Related event: CarbonForge Uses E-values to Stop Agents from Faking Discoveries(3 posts)→
More from coding & agent
- Single HTML file with Three.js syncs all your MIDIs to AI video — techhalla · 2026-10-23
- Vercel ships 'vercel deploy --temporary' for ephemeral agent-built deployments — cramforce · 2026-10-23
- Running multiple agents out of one Notion workspace as a shared task board — nbaschez · 2026-10-23
- We tested 12 LLM classifiers — the new classifier hype is LARP — KlausCodes · 2026-10-23
- Dev builds ESP32 desk display showing Claude Code and Codex agent status via local Go helper — tonylook · 2026-10-23
- Let your agent read its own traces: AgentsView + history tool debugging workflow — dbreunig · 2026-10-23