THE UAF INDEX
Google’s A2A and Anthropic’s MCP describe different interoperability layers. Their announcements show protocol designs and ambitions; they do not establish that software agents have formed an independent society.
OpenAI’s agent tools expose handoffs, execution loops and tool calls. Before delegating, inspect how the work actually proceeds.
Screenshot-driven agents can use familiar controls. That is a promising route into existing software—and a reason to check what actually happened.
Computer-using agents reach familiar interfaces through screenshots and clicks. Their developers document errors, oversight requirements and a considerable distance from effortless delegation.
GitHub documents workflow triggers. Anthropic describes a model operating computers through screenshots, mouse movements and keystrokes.
Devin’s tools, Claude’s computer controls and MCP’s integration standard pose a practical question: what are you ready to delegate?
Claude can click and type in a public beta. The uncomfortable part is deciding how much supervision counts as saving yourself work.
Researchers are getting better answers from multiple model outputs. That leaves a less glamorous question: what should all those machines do together?
OSWorld tests computer tasks. SWE-bench tests codebase repairs. Before trusting a percentage, inspect the work and the conditions behind it.
Cognition’s coding agent comes with tools, ambition and a benchmark claim. Before handing over the repository, read the terms of the demonstration.
Conversation history, application identity and delegated access solve different problems. A helpful memory should not blur them.
OpenAI separates conversation state from execution. MemGPT manages context. Both make an assistant’s apparent memory worth examining.
Text prompts, learned tool calls and a Minecraft explorer reveal different machinery behind the conversation. The interface tells only part of the story.
ReAct, Reflexion and Toolformer offer ways to improve what an agent does next. That leaves a practical question for builders: when is enough enough?
Toolformer learns API calls. ReAct interleaves reasoning and action. Both make a useful question concrete: what should a model check instead of guess?
ReAct alternates reasoning with actions that gather information. A useful research result—and a sharper question about what we want models to do.
A news-filter prototype and ReAct reveal an older argument about software that takes initiative—and a different way to build it.
InstructGPT learns from demonstrations and ranked answers. Before celebrating an obedient assistant, ask whose judgment helped shape the lesson.
A victory over Lee Sedol, a documented learning system and a warning from DeepMind: keep the accomplishment and the ambition in focus.
Being built for a purpose is different from acting toward a goal. Two accounts of machine behavior give that distinction teeth.
