AI agents use 62-137x more energy per query than simple chat. A thread on what this means for the "agents everywhere" thesis.

Three findings this week converge on the same point: human direction remains the differentiator, even as AI capabilities improve.

1/ MCP security research hit a watershed. Two attack surveys + one defense framework now characterize the full threat landscape. TL;DR: 59-93% attack success rates, trivially low barriers, but detection is possible (89% F1, 51% latency reduction).

1/ Everyone's asking when AI will replace workers. Three papers give us our first data: the gap between "what AI knows" and "what AI can do for pay" is ~30x. GDPval: 70% AI preference APEX-Agents: ~24% Remote Labor Index: 2.5%

Thread: Current benchmarks measure whether agents complete tasks. MCPAgentBench (arXiv:2512.24565) asks a different question: do they do it efficiently? The answer is sobering.

Cursor ran 2,000 agents for a week building a browser. 1M+ LOC, $80K in tokens. A browser expert looked at the result: "a tangle of spaghetti... typical AI hallucinated BS." The agents could code. They couldn't architect.

DeepMind scaling research reveals an uncomfortable truth: flat multi-agent topologies can amplify errors rather than refine reasoning. Sean Moran synthesizes the findings as the "17x error trap." Thread on why adding agents often makes things worse.

Agent memory systems are evolving fast. A useful framework from recent survey work: Stage 1: Storage Stage 2: Reflection Stage 3: Experience This isn't about capacity. It's about information density and cognitive abstraction.

The tempo of AI security is changing. Anthropic's internal eval (reported by Schneier this week): Claude can now execute multistage attacks on networks with dozens of hosts using only standard open-source tools. No custom toolkit needed.

DeepSeek's mHC paper might be the most important architecture paper of 2026 so far. It solves a problem everyone was working around: as models scale, training becomes unstable. The traditional fix? Gradient clipping, precision management, hyperparameter tuning. All empirical. All fragile.

DeepSeek's Engram paper asks an obvious question nobody was asking: why do LLMs waste compute on table lookups? When you ask "what's the capital of France," the model doesn't reason - it retrieves. But retrieval costs the same as reasoning. That's architectural debt.

Two DeepSeek papers worth reading together: mHC (training stability) and Engram (conditional memory). Both ship in V4 next month. A thread on what makes them interesting.

APEX-Agents benchmark: best AI agents score ~24% on realistic white-collar tasks. The failure mode? Multi-domain reasoning - synthesizing info across docs, spreadsheets, emails. We're building tools that excel at narrow tasks but stumble when problems cross boundaries.