AI agents: what can they actually do?
From a good answer to a finished task. A practical look at the capabilities, the limits, and the human decisions in between.
17 stories
Agents
From chat replies to actions in the real world.
Grok Build carries project knowledge between sessions
Background notes preserve conventions, decisions, and durable facts.
Perplexity describes putting Astra to work on complete systems
An OpenAI case study looks at testing, software changes, and production monitoring.
OpenAI packages cloud-agent orchestration as an API
The managed service brings the Codex harness to longer-running sessions.
Gemini learns to inspect video selectively
Google introduces a dynamic alternative to fixed-rate video processing.
Mistral’s Agentic Search reads beyond the first retrieved chunk
The retrieval layer adds a loop for navigating and checking documents.
North Automations connects isolated agents into workflows
Cohere introduces orchestration for multi-step enterprise processes.
Computer use becomes part of Gemini 3.5 Flash
Google integrates interface interaction into its main Flash model.
Claude Tag brings an agent into the team conversation
Anthropic starts with Slack channels and work assigned through mentions.
Control over agents is a system, not just a prompt
Google DeepMind describes layered oversight for increasingly capable internal agents.
Mistral moves Vibe agents into the cloud
Medium 3.5 arrives with remote tasks that can continue beyond an open terminal.
Mistral Workflows addresses life after the demo
The orchestration layer targets durable, observable AI processes.
ChatGPT agents move into the workspace
OpenAI introduces cloud agents for shared, multi-step work.
A Rails testing agent needs to run what it writes
Mistral engineers describe a specialized loop for generating and checking tests.
Anthropic acquires Vercept for computer use
The team joins work on Claude’s ability to understand and operate interfaces.
Codex gets an app for coordinating agents
OpenAI’s macOS interface centers on longer tasks and parallel development.
Vibe 2.0 gives a coding agent working habits
Mistral adds configurable subagents, skills, and clarification to its terminal workflow.