The archive
Systems that plan, use tools and act on your behalf.
DeepMind says a new set of Cheminformatics Skills lets an agent in Google Antigravity curate literature compounds, train potency models and propose macrocycles in about an hour. No wet-lab validation was reported.
5 min read
ServiceNow will route workflow actions into ChatGPT and put OpenAI models behind development, security and voice features. Today's shipping piece is an MCP plugin; the rest lands in 2027.
5 min read
In an OpenAI customer story, Greptile says it uses GPT-6 Astra's computer use to click through pull requests like a user would, plus Codex and an MCP knowledge base to cut reverts.
5 min read
AWS's monthly roundup adds OpenAI-powered Managed Agents in preview, a leaner AgentCore runtime, an open source Strands harness and a wave of new frontier models on Bedrock.
5 min read
Postman and AWS detail the architecture behind Agent Mode, including dynamic tool selection past a 40-tool accuracy cliff, schema-aware reads and tiered prompt caching on Amazon Bedrock.
6 min read
OpenAI's customer story credits GPT-6.1 Sol and Codex for a 76x cost drop in Asana's browser agent. The underlying numbers suggest prompt caching and history management did most of the work.
6 min read
Sophos says agents built through OpenAI's Daybreak program now resolve 52% of managed detection and response cases end-to-end, cutting average investigation time by 96%.
5 min read
An astrophysicist used a team of Claude agents to merge decades of UV space-telescope data and predict the third of the sky never observed. The human still caught the bug the agent reviewers missed.
5 min read
OpenAI's customer story on Pollo AI describes a two-tier routing setup across GPT-5.6, GPT-6 Astra and GPT-Image-2.5, and claims users spend over 50% less time switching models.
5 min read
Harvard's Matthew Schwartz says he stopped asking Claude to act like a scientist and built a harness around what it does well. The results span 18 fields — and reveal where agents still fail.
6 min read
In Project Swap, 201 Anthropic staff sent Claude agents to barter books on a trading floor. The bargaining was fine; the preference elicitation was where most of the value leaked away.
6 min read
AWS published a reference architecture using Bedrock AgentCore Gateway and MCP to let one agent query data held in separate line-of-business accounts. The trade-offs sit in the authorization layer.
6 min read