Business
Google DeepMind published a 42-page technical report titled “Intelligent AI Delegation” with a straightforward conclusion: almost all systems sold as AI agents today are actually well-packaged task performers. The document, released in February 2026, argues that delegating is not dividing a task and distributing it to different models. Delegating involves transferring authority, responsibility and trust dynamically, something that current systems do not do.
The problem is structural. When an agent receives a goal today, it decomposes the task, calls tools, and returns a result. DeepMind classifies this as automation, not delegation. The difference matters because, without a formal accountability structure, any chain failure brings down the entire system. The report organizes real delegation into five pillars: dynamic assessment (measuring risk, cost and reversibility before passing the task), adaptive execution (reassigning tasks during execution when something goes wrong), structural transparency (the agent needs to prove what it did, not just report), trust calibration (humans trust AI too much, and agents trust each other too much) and systemic resilience (if everyone delegates to the same “best” model, a failure in it brings everything down).
The agent market is expanding. Companies sell multi-agent solutions for service, sales, operations and code. But most operate on a linear logic: agent A passes to agent B, who passes to agent C. No formal permissions, no verification, no fault assignment. DeepMind does not propose that agents stop existing. He proposes that delegation becomes a protocol, not a prompt. The document’s most straightforward sentence sums up the position: “Automation isn’t just about what AI can do. It’s about what AI should do.”
What changes for the sector
The distinction between automation and delegation has practical consequences for companies that set up multi-agent systems. Without an accountability structure, failures accumulate without anyone being able to trace their origin. The report points out that we are moving from “prompt engineering” to “agent engineering” and now to “delegation engineering”. Companies that solve intelligent delegation first will build truly autonomous systems. The others will deliver fragile demonstrations.
Source: Google DeepMind, “Intelligent AI Delegation”, 42-page technical report, February 2026.
