Anthropic engineer on building self-improving AI systems: loops and graphs explained
goyalshaliniuk · x · 2026-09-03
An Anthropic engineer explains in a 41-minute talk how the company uses AI internally and lays out a progression: Prompts → Agents → Loops → Graphs.
- Loops let a unit of work check and correct itself.
- Graphs decide which units should exist, connect them, and turn successful outcomes into rules for the next run — which she argues is what real self-improvement in AI systems looks like.
Her core claim: to become a top 1% AI engineer, don't just build agents — build systems that improve themselves, and the graph layer is the part most people still aren't building.
More from coding & agent
- 1Password on safely using credentials with agents; Proton Pass and Bitwarden as open-source alternatives — uwukko · 2026-09-03
- OpenAI Codex jumps from 6M to 25M active users in 7 weeks, over 4x growth — FinanceYF5 · 2026-09-03
- ARBR open-sources an MIT-licensed routing and governance layer for multi-provider LLM traffic — Genie-Tickle-007 · 2026-09-03
- Anthropic shares 6 official tips to cut Claude Code token costs — addyosmani · 2026-09-03
- Coolify MCP Server Lets AI Agents Deploy and Monitor Self-Hosted Apps via Natural Language — modelcontextprotocol · 2026-09-03
- GOV.UK Ships an MCP Server for Search, Content Retrieval and Postcode Lookup — modelcontextprotocol · 2026-09-03