loading…
Agenticness is bounded runtime decision-making, not merely tool use
A system becomes agentic when a model is allowed to decide part of the execution path at runtime. The model is not just generating content inside a fixed step; it is selecting actions, interpreting observations, and deciding whether the goal has been reached.
Consider two systems that both use a search tool:
Both systems may use the same model and tools. The difference is where decision authority lives.
The loop repeats until a stop condition is reached. Different implementations may call this planning, reasoning-and-acting, orchestration, or a state machine with model-directed transitions. The naming matters less than the control pattern.
| Capability | Why it matters |
|---|---|
| Goal | Defines an outcome that may require multiple steps rather than one response |
| Action selection | Allows the model to choose among tools, questions, or sub-tasks |
| Observation | Feeds tool results and environmental changes back into the next decision |
| State | Tracks completed work, unresolved questions, and intermediate artifacts |
| Iteration | Supports retrying, revising, or replanning when evidence changes |
| Stop conditions | Prevent endless loops and define success, failure, escalation, or budget exhaustion |
Agenticness is not a binary label. A system may allow the model to choose a tool but not authorize write actions. Another may let it create a plan but require a human to approve execution. A background coding agent may read files, edit a branch, and run tests, but be unable to merge to production.
| Level | Model authority | Example |
|---|---|---|
| Assisted | Generates recommendations; human executes | Draft a database migration plan |
| Bounded | Acts within approved tools, data scopes, and budgets | Create a support ticket after validating required fields |
| Supervised | Plans and executes low-risk steps; requests approval for consequential actions | Prepare a refund, but require approval before issuing it |
| High autonomy | Runs long tasks with broad action authority and periodic review | Maintain a test suite in an isolated repository environment |
The right autonomy level depends on reversibility, financial impact, data sensitivity, and how well success can be verified.
An agent should not receive every available tool and an open-ended instruction to "figure it out." Good architecture defines:
These are building blocks or patterns. A system is agentic when the model has meaningful runtime discretion over the path.