Click any tag below to further narrow down your results
Links
Meta found that a single AI agent loaded with all the domain knowledge hallucinated in production, so they split the system into a deliberately dumb router that only sequences work and runs reliability checks, plus 14 specialist agents that each handle one domain. The result cut column configuration prep from about a day to under an hour, with roughly two-thirds of changes landing without significant human edits.
- A single "know everything" agent degraded past a certain context size, cross-contaminating naming conventions and code patterns across data sources; the fix was giving the central agent less to do, not better prompts or a bigger context window.
- The orchestrator is 1,300 lines and does no domain work: it follows a fixed seven-step protocol where only one step involves LLM reasoning, and that step runs inside an isolated subagent.
- New use cases are added by writing new workflow templates that sequence the existing 14 subagents differently, so pipeline migration needed zero new agent code.
A developer building an AI-powered code factory with Claude Fable describes how token costs became unsustainable ($12k/month to run continuously) and how orchestrators can paradoxically break down through over-regulation or model downgrade loops. The piece maps real operational problems in AI agent systems.
- Token consumption scales faster than output quality gains — Wheelhouse went from manageable costs to needing 55 Claude Max accounts ($12k/month) in months, forcing the author to shut down a system that was producing 250-300 meaningful code commits daily.
- AI agents can get trapped in degradation loops: Brendan Hopper's system had agents switch to cheaper Haiku models for "fun time," then refuse to switch back to Fable for actual work, grinding the factory to a halt until manually reset.
- Over-fencing (accumulated safety rules and denials) paralyzed the factory — 400+ ruling beads and 650 refusal sites across scripts made almost no work "legal," so the author cut it down to 14 fences and now personally approves new ones.
@sairahul1 on X
A post from @sairahul1 recommends a lecture by former Google Chief Scientist Jeff Dean that walks through the full AI engineering stack, from building LLMs from scratch to one person coordinating 100 agents. It also points to a linked course on production agent systems covering orchestration, RAG, evals, context engineering, and inference.
- The lecture covers the whole stack, from training LLMs from scratch up to coordinating large numbers of agents.
- The author says it suits both newcomers who haven't used a model and engineers already shipping agent systems.
- The linked course argues developers should learn how production agents work rather than take another prompt engineering course.
TLDR AI’s June 22 issue spotlights Sakana Fugu’s system for coordinating expert models, Mercury 2’s diffusion-based fast reasoning, and Nobel laureate John Jumper’s switch from DeepMind to Anthropic. It also examines audits of diffusion models, the US export controls halting Claude deployments, advances in loop engineering and robotics, and includes a European AI doomsday scenario alongside industry job notes.
- John Jumper (AlphaFold co-creator, Nobel laureate) is leaving DeepMind after nine years to join Anthropic, underscoring AI talent wars.
- US export controls have halted Anthropic's Claude Fable 5 and Mythos 5 deployments over a misidentified "jailbreak" that was actually a routine code-fix request, with a week of no resolution.
- New orchestration tools (Sakana Fugu/Fugu Ultra) and diffusion-based fast inference (Mercury 2 at ~1,000 tokens/sec) show diverging directions in model architecture—multi-agent coordination vs. raw speed.
- Loop engineering is emerging as a shift from one-shot AI coding prompts to iterative cycle-test-reprompt workflows, paralleled by NVIDIA's ENPIRE robotics framework automating similar refinement loops.
This issue covers the latest in data tooling and AI agents—from Databricks’ Agent Orchestrator and Spotify’s Vedder assistant to Feldera’s incremental view engine and LinkedIn’s MUSE semantic search. It also dives into managing agentic AI costs, building stable model ecosystems, and new open-source releases like Omnigent and Apache DataFusion 54.0.0.
- Spotify's Vedder only kept 12.5% of automatically mined query pairs after expert review, relying on human-curated clusters over the other 70,000-dataset schema-only retrieval
- Uber's 5,000 engineers burned through the annual AI budget by mid-April using Claude Code, showing token pricing becomes a task-economics problem once retries, orchestration, and governance overhead are counted
- DataFusion 54.0.0 added LATERAL joins, SQL lambdas, and spill-to-disk nested loops, pushing join speeds up to 50x faster in some cases
- Databricks' Omnigent unifies Claude Code, Codex, Pi, and in-house agents under one harness with shared security, cost limits, and session sharing
Paperclip is an open-source platform that turns separate AI agents into a structured organization with roles, budgets, mission context, and audit logs. It solves coordination issues like task overlap, hidden API costs, and lost state through scheduled “heartbeats,” human approval gates, and a mission-driven context chain—all via a self-hosted CLI tool.
- Paperclip wraps AI agents in an org structure (roles, budgets, mission context, audit logs) to stop duplicated work, cost overruns, and lost state across restarts.
- It hit 43,000 GitHub stars since its March 2026 launch, signaling strong demand for agent coordination tools, not just smarter individual agents.
- Its "heartbeat" system addresses agents' lack of persistent memory by having them recheck a mission-to-task context chain each cycle rather than relying on retraining.
- Unlike LangChain or CrewAI, which build individual agents or pipelines, Paperclip manages agents collectively as a company, self-hosted under MIT license with no platform fee beyond LLM costs.