In May 2026, Anthropic described a technique it called “dreaming” — letting an autonomous agent review its own past behaviour between sessions, spot patterns, and adjust how it works next time. It’s part of a broader industry push toward agents that handle long-running workflows in areas like coding, finance and legal work, rather than one-shot tasks. That shift changes what you have to build around an agent.
One-shot vs. long-running
A one-shot agent takes a request and returns a result; if it’s wrong, you try again. A long-running agent holds a goal across hours or days, accumulates state, and takes many actions before anyone checks the outcome. The failure modes are different: drift, compounding small errors, and actions taken on stale assumptions.
What “learning between sessions” needs from you
- A durable record of what the agent did and why, that both the agent and a human can review.
- Checkpoints where a person confirms direction before the agent commits to the next phase.
- Clear boundaries on what the agent can change without approval, especially as it gets more autonomous.
The trust curve
Agents that improve over time are more useful and harder to reason about — the thing making decisions this week isn’t quite the thing you evaluated last month. Treat capability growth as a reason to strengthen oversight, not relax it.
Summary: Long-running, self-improving agents are a real step forward in usefulness. The operational discipline they require is a step up too, and it doesn’t come in the box.
#AI #AgenticAI #Automation #EnterpriseAI #SazkoSolutions