Agents.
Four ways to look one in the eye — a benchmark, a harness, a panel, and a trace.
The long jobs are where they fold.
OSWorld 2.0 stops testing agents on toy clicks and hands them 108 real workflows — the kind a person finishes in about an hour and a half. The best model, Claude Opus 4.8 at full thinking, finishes one in five. Short benchmarks flattered everyone; stretch the horizon to ~318 tool calls and the scores collapse.
fully solved
Read the whole agent, line by line.
xAI open-sourced Grok Build — its terminal coding agent, the full Rust harness, under Apache 2.0. That's the agent loop, the tools that read and edit your code, the TUI, and the whole extension surface — MCP, skills, plugins, hooks, headless, ACP. You can compile it, point it at your own inference, and audit exactly what it does before it touches a repo.
The panel that only agrees with itself.
Spin up five sub-agents on the same model and the same context and you get one opinion wearing five name tags. The Council skill names the trap: correlated votes reduce error by roughly nothing, yet the apparent unanimity inflates your confidence. The fix is mandates, not costumes: give each seat a different job and different evidence, then treat disagreement as the only reading worth trusting.
N seats ≈ ONE sample, quoted N times.Persona is decoration. The mandate is the whole thing.
Follow the slow request, span by span.
PostHog's distributed tracing is in beta, and it takes your existing OpenTelemetry exporters — no new SDK to wire in. Errors tell you something broke; logs show scattered events; a trace draws the causality chain, with timing. Because the spans sit next to product events in the warehouse, you can pivot from a 3.2s checkout to the exact N+1 query, then jump straight to that user's session replay.
| Span | Service | Duration | Time |
|---|---|---|---|
| checkout | web | 3.2s | |
| inventory · N+1 | inventory | 2.8s | |
| charge | payment | 0.3s | |
| confirm | 0.1s |
That's today.
One thread ran through the picks: what an AI agent is actually doing when you look closely — how far it gets, what's inside its harness, when a panel of them is fooling you, and how to trace it in production.