Agentic AI
Also: AI agents · autonomous agents
Software that uses a language model to decide a sequence of actions and then carries them out against real systems, rather than only returning text.
The distinguishing property is not intelligence but consequence. An agent calls tools, writes files, sends requests and spends money. Everything that follows — permissions, budgets, audit trails — follows from that and not from how the model reasons.
Orchestration is the live design question
A central planner decides which agent does what. Research on market-based alternatives found the weakness plainly: in a centralised allocator built on a language model, inserting a single preference nearly doubled the favoured agent's share of tasks. One line in a context window moved the distribution of work that far.
That is a structural argument, not an efficiency one. A planner whose routing policy can be rewritten by anything reaching its context is a component with a security property, and a bidding mechanism moves the decision into a rule that does not read prose.
What has to exist before an agent may act
- Attribution: which agent, acting for whom, on what instruction.
- A limit enforced outside the agent rather than inside its prompt.
- A record that survives the agent and can be produced at audit.