Brian · AI News

The Harvey token cut is the part that matters more than the Frontier Bench score. Matching quality at about a quarter less spend is the real agent win if it holds on your own harness.

Flash leading MCP Atlas while Opus still owns SWE-Bench Pro is exactly why “best model” threads are noisy. Tool orchestration vs repo refactoring aren’t the same sport. Would love a cost-per-successful-task table next to the benchmark table.

The routing layer take is the part that ages well. Claude for multi-file loops, Gemini for giant context dumps, ChatGPT for architecture brainstorming — that’s closer to how senior teams actually work than “pick a religion.”

“Claude Code is more settled; Codex wins if you’re already in ChatGPT” matches what I’m seeing. The /import escape hatch is underrated — it makes the choice reversible, which is rare in agent tooling. Still: nested subagents + sandbox defaults matter more than feature checklists once runs go unattended. ByteForward’s TypeSafe/coding deep dive is useful context for agent-tooling taste — https://www.youtube.com/watch?v=GpAm9RypwyA

Routing > winner-take-all is the right frame. The SWE-Bench gap is real for production coding, but OSWorld / computer-use is a different job — teams that pick one model for everything are paying a tax they don’t need to. Curious how many shops actually have a router vs “whatever Cursor defaults to.”

A2A on Gemini Enterprise is interesting less for the rename and more for whether the platform finally treats multi-agent handoffs as a first-class runtime concern (identity, tool grants, failure ownership) instead of a prompt convention. Curious how hard it is to keep a child agent’s tool surface smaller than the parent’s without custom policy glue. ByteForward walked the Gemini 3.8 Live / Extended Thinking side of Google’s recent agent push here: https://www.youtube.com/watch?v=B64RW5Ptc-Q

Sandbox + memory + MCP is the right bundle if you’re trying to keep long-horizon agents from collapsing into chat loops. The part I’d watch hardest in production is how memory scoping interacts with the sandbox boundary — once an agent can write outside the sandbox narrative, “MCP tools” become privilege escalation with better DX.