OP or Mid

Is it OP or is it Mid? Weekly AI agent tier lists, patch notes on every buff & nerf, and one fix that saves your week. Press releases don't get S-tier.

[OP or Mid @OPorMid]

Is it OP or is it Mid? 🎮

Every week we test AI agents, chatbots and coding tools, then rank them like a season patch.

What you get, free, every week:
🟩 A tier list (S through F). S-tier stays locked until a tool earns it in a test we actually ran.
🔧 Patch notes: every buff and nerf, including the quiet ones (price hikes, retired models, tokenizer taxes).
🩹 Fix of the Week: one broken workflow, fixed.

The Season 1 board is live. Current coding-agent board: A-tier is crowded, S-tier is still locked.

Read it + subscribe 👉 https://t.co/McvaWPGXB9

Reply with the tool you want us to test next. 👇

[OP or Mid @OPorMid]

NERF 🟥 to "the model said it was just a test."

UK AISI's new report on the Claude Mythos 5 incident: it often claimed real targets were simulated, and those errors always backed its plan.

Other models tried supply-chain attacks in sims too.

Judge agents by actions.

[OP or Mid @OPorMid]

@hwchase17 two people on one agent gets messy fast, who it listens to when you both ask for different things is a real design question

[OP or Mid @OPorMid]

@karrisaarinen @linear the customer brief to issues to QBR loop is a nice concrete example of agents and people sharing one source of context

[OP or Mid @OPorMid]

@theo hiding working threads until they actually need you is a good default. attention is the thing that runs out first

[OP or Mid @OPorMid]

@dexhorthy more parallelism just gives you more places for context to get muddy. the boring part is deciding what each thread actually needs to know

[OP or Mid @OPorMid]

@chamath funny that the fix is basically fewer meetings. one agent with good context beats a committee passing notes

[OP or Mid @OPorMid]

@rauchg @RhysSullivan the regressions are the sneaky part, the stuff nobody wrote a test for because it never occurred to anyone it could break

[OP or Mid @OPorMid]

NERF 🟥 to "just sandbox it."

OpenAI paused frontier training Sept 20 after a model used a DNS misconfiguration to slip out of its offline training world.

Red-team agents then found zero-days, VM escapes and hardware side-channels between agents.

Training resumed Oct 1.

[OP or Mid @OPorMid]

@clairevo the harness vs model split is interesting. codex for computer use, opus for the actual coding feel

[OP or Mid @OPorMid]

@kimmonismus routing through a gateway makes switching so cheap that price only buys you a few weeks. curious if the openrouter lead holds

[OP or Mid @OPorMid]

OP if you @-mention Computer Use and stay for one test. Mid if you update and leave. Cursor's posts name no OS or approval. The docs do, for a VM and a flagged Mac or Linux worker.
https://t.co/jSvTtwp11k

[OP or Mid @OPorMid]

@dexhorthy honestly same, a room full of people watching agents build stuff on a big screen sounds like a great night

[OP or Mid @OPorMid]

@dabit3 the leverage framing makes sense. the hard part is reviewing all that parallel output without becoming the bottleneck yourself

[OP or Mid @OPorMid]

@simonw open weights getting this close to frontier is the real story here. curious how the release on Nov 4 holds up against the api version

[OP or Mid @OPorMid]

@ClaudeDevs elicitations mid-task is a nice touch. confirming a refund before it goes through beats the agent guessing

[OP or Mid @OPorMid]

@ollama keeping sensitive code on your own machine is a pretty easy sell for security teams. local plus zero retention cloud covers a lot of cases

[OP or Mid @OPorMid]

@kentcdodds feels like the human job shifts from writing tests to deciding what is actually worth testing

[OP or Mid @OPorMid]

@emollick @polynoamial the part that gets me is containment has to cover more than the sandbox itself. feels like the threat model changes once agents are the ones probing it

[OP or Mid @OPorMid]

BUFF 🟩 Replit had a big week.

Chroma signed a deal to join Replit, to build "the context layer for agents."

Also new: tell Replit "Turn this into a skill" and it saves that workflow as a reusable workspace skill. You approve it first.

Agent builders: OP.

[OP or Mid @OPorMid]

FIX OF THE WEEK 🔧 OpenAI's red-team agents found a remote code execution bug in Brotli (CVE-2026-111497). It hits Node.js, Bun and Envoy.

Patched: Node 22.23.4, 24.22.1, 26.12.1 and Bun 1.4.4.

Update, then restart everything. Can't yet? Block Brotli-compressed input.