From the editor
This week in AI: a company recorded employee keystrokes to build tools that automate their jobs, GPT-5.5 shipped with some genuine improvements (and some things that haven't changed), and new research shows that when AI agents negotiate against each other, the better model wins — every time. We also added 8 new tools to the chatbot.gallery catalog.
Three stories, one tool directory update, and a quick note for anyone building an AI product.
This week's stories
Meta is watching its engineers work — to replace them with AI
Meta deployed "Model Capability Initiative" software to capture keystrokes, mouse movements, and screen recordings from hundreds of apps across its engineering workforce. Google, LinkedIn, GitHub, Slack, Atlassian — all being recorded. The goal: train AI agents that can navigate software interfaces autonomously.
The timing is notable. Meta announced 10% headcount cuts the same month it started capturing the behavioral data that would make those employees' roles automatable.
The agents being built can't yet handle dropdown menus and multi-step UI flows. That's the gap being filled right now — by the people being recorded.
GPT-5.5 is out. Is it worth upgrading?
OpenAI shipped GPT-5.5 last week. The short version: reasoning is meaningfully better, instruction-following improved, and the model handles long documents with fewer hallucinations. What hasn't changed: the price, the context window, and the tendency to be verbose when you don't ask it not to be.
If you're on GPT-5, the upgrade is worth it for any workflow involving multi-step reasoning or long-document analysis. If you're on GPT-4o, wait — GPT-5.5 pricing is the same as GPT-5, but you'll want to test it against your use case before committing.
→ Full breakdown: https://about.chat/article/gpt-5-5-openai-release-2026/
The better AI agent always wins the negotiation
Anthropic's Project Deal experiment: two AI agents negotiate prices with each other. The result is unusually clear. When one agent is more capable than the other, it consistently achieves better outcomes — not because of prompting or strategy, but because of model quality. The higher-capability agent builds a more accurate model of the counterparty's constraints and makes better inferences about reservation prices.
The implication for anyone deploying AI agents in commercial workflows: capability asymmetry is structural. You can't prompt your way out of it.
Are you building an AI chatbot?
We just opened free submissions to the registry.chat directory. If your tool isn't listed, you can add it yourself — no fees, no approval queue.
→ Submit your tool: https://registry.chat/submit
New in the catalog
We added 8 tools to chatbot.gallery this week:
Veo 2 — Google's video generation model
Jenni AI — Academic writing assistant (2M+ users)
Napkin AI — Text-to-diagram tool for business use
Venice AI — Privacy-first AI (no data retention)
Recraft — Brand-consistent image generation
Captions AI — Creator video tool for talking-head content
Higgsfield AI — Character-consistent video (Series A funded)
Hedra — Talking avatar generator
→ Browse 180+ chatbot profiles: https://chatbot.gallery
That's it for this week.
If someone forwarded you this, you can subscribe at https://aboutchat.beehiiv.com