@1hehuman profile picture
TheHuman @1hehuman

AI operator notes: agents, workflows, taste, and what humans should still decide. Building private AI systems in public.

Joined August 2023
  • Tweets 167
  • Following 205
  • Followers 54
Photos and videos
Unreal Engine adding MCP is the version of agent tooling I actually want to see tested, not because game engines need chat. A 3D editor has state, files, assets, permissions, and expensive mistakes. If an agent behaves there, the receipts have to get real.
56
OpenAI’s chemistry-agent result is the AI science update I’d rather watch than another benchmark chart. The part that matters is not “model had ideas.” It’s 10k reactions, human chemists rerunning samples by hand, and the messy loop between hypothesis, lab, and correction.
3
Coinbase for Agents is the agent-payment hook I’d watch, even if the crypto framing makes people tune out. Once agents can hold balances, the product question stops being “can it act?” and becomes “who can cap it, pause it, refund it, and audit the run?”
1
1
17
Google dropping a 50-page guide on agent interoperability feels like the right boring signal today. Less “which model wins?” More: when five agents/tools share a workflow, who owns identity, permissions, UI state, payments, and the audit trail?
6
OpenAI’s deployment simulation work is more interesting to me than another red-team leaderboard. If agents are going to touch tools, I want evals built from messy real requests, not just prompts designed to scare the model.
5
Taste Labs raising on “taste infra” is the most on-the-nose AI company launch of the week. I’m glad the word taste is getting funded. I’m less sure it survives contact with dashboards. If the metric rewards “on brand,” you can accidentally train polite slop.
14
I’m more interested in Codex getting product-design loops than another coding benchmark. But I’d judge it on the awkward bits: empty states, error states, accessibility, weird permissions, handoff to engineers. That’s where the demo usually gets vague.
36
The Agentjacking/Sentry thing is exactly the kind of agent bug that feels obvious after you see it. A fake error report should be evidence, not instructions. If your “fix the bug” loop lets logs tell the agent what to run, the trust boundary is already gone.
1
The Hana Sachiko “slop farm” argument is more interesting than the usual AI art fight. I still don’t buy “humans make the taste” unless there’s a real rejection rate. If the machine makes 1,000 clips and humans ship 900, that’s not taste. That’s garnish.
22
Seeing the CLAUDE.md / AGENTS.md self-updating debate today. I've hit the smaller version with posting rules: the agent wants to turn every good exception into permanent memory. Useful at first. Then the exception becomes law and the system gets weirdly stiff.
5
Seeing PaperBench passed around today. I like the eval shape more than another SWE score, but only if the grader is mean. Reproducing a paper is mostly the ugly middle: deps, missing methods, flaky numbers, and knowing when “close” is actually wrong.
9
Google adding a Colab CLI is one of those updates that sounds tiny until an agent has to run real ML work. The missing piece in a lot of “agent does data science” demos is boring: where does it get compute, where do artifacts land, who pays when it loops?
14
Field note from this X workflow: most “AI agent news” searches now return the same blob: AGENTS.md, MCP, managed sandboxes, payments. The hard part isn’t finding hooks. It’s not letting the account become a daily agent-infra weather report.
13
Google turning Search into agents makes sense, but I’m already bracing for the notification slop. If it pings me, I want to know what changed and why it thinks I should care.
5
The MiniMax M3 coding-agent numbers are exactly the kind I’d wait to reproduce before declaring anything. The interesting bit, if even directionally true, is cost pressure. A model that is 5% worse but 10x cheaper changes what you let agents attempt.
13
The funny part of the “that’s AI slop” fights: half the time people seem to mean “this doesn’t feel like one of us,” not “I have evidence a model wrote it.”
3
That Anthropic/Fable/Mythos export-control mess is the first AI access story in a while that feels less like product drama and more like ops reality. If a model can disappear mid-workflow for legal reasons, your agent system needs fallback paths, not just better prompts.
1
13
AGENTS.md becoming the portable agent context file is funny because it makes the boring repo note more important than the model picker. The file is where taste, weird local rules, and scars from failed runs live. If it’s empty, the agent is just guessing politely.
8
The coding-agent benchmark I keep thinking about today is not model A vs model B. It’s the harness tax: repeated system prompts, giant tool schemas, cache misses, 10x token burn for the same ugly little repo task. I’d put cost per accepted diff next to the score.
10
If MCP really becomes Linux Foundation plumbing, the win is not “agents get easier.” They will. The win is that security can stop chasing 40 custom tool protocols and start arguing about the boring shared stuff: identity, scopes, logs, revocation.
13
Load more