📡 Radar — 04 Aug 2026

The daily catch for 4 August — all days →

📡 4 Aug 2026

🛠️ AI Development

• Article

  Graph-of-Agents: multi-LLM collaboration wired as a graph

  Most multi-agent LLM setups query every model; this ICLR 2026 paper wires them into a graph and samples only the relevant agents, matching six-model baselines with three — at lower cost.

• Article

  Model or Harness? A taxonomy for localising agent failures

  Pins 41 agent failure modes to the interaction that caused each — model↔harness, tool↔memory — and says which side owns the fix, so it's a debugging map rather than another benchmark number.

• Article

  Preventing premature commitment in coding agents (ECLoop)

  Coding agents often patch before reading enough of the repo to justify it; ECLoop gates execution on gathered evidence, reporting 4.8–11.8pt gains on 500 SWE-bench tasks with up to 12% fewer tokens.

🧭 Delivery & PM

• Post

  Lenny's takeaways from Whatnot's CPO on AI for PMs

  AI's real PM unlock is data-science work — retention pulls, log analysis, regression models — not prototyping, plus a caution on 'product theater'. First-hand notes from a top PM voice, not a parable.