AI digest: The agent takeover accelerates
OpenAI declares chat dead whilst everyone rushes to build autonomous agents that actually do things.
The shift from chatbots to agents is happening faster than expected. Companies are moving beyond conversation to systems that take action.
OpenAI bins ChatGPT for agent superapp
OpenAI says “chat is dead” and plans to rebuild ChatGPT as a full agent platform bundling coding tools, booking services, and partner apps like Canva. This makes sense given how limiting the current chat interface feels when you want AI to actually do tasks rather than just talk about them.
UIUC builds proper retrieval agent with RL
Researchers released Harness-1, a 20B parameter retrieval agent trained with reinforcement learning that maintains its own search state and evidence graphs. It beats other open agents by 11 points on benchmarks, showing that purpose-built retrieval systems crush general chatbots at information tasks. The weights are public, which is refreshing.
Meta’s $200 agent puts skin in the game
Meta is testing “Hatch”, an AI agent that costs up to $200 monthly and builds working tools from simple descriptions. This marks Meta’s first paid AI product after years of free models. Charging real money forces them to deliver actual utility rather than impressive demos.
DeepSeek dominates US enterprise spending
DeepSeek topped corporate AI spending in June as US companies chase cheaper alternatives to OpenAI and Anthropic. The cost pressure is real, but sending sensitive data to Chinese models feels like the kind of decision finance loves and security teams hate.
Perplexity lets models write their own search code
Perplexity launched “Search as Code” where AI models write Python search routines instead of using fixed APIs. This is clever because it lets agents adapt their search strategy to each query rather than being stuck with predetermined workflows.