CORE GUIDE
Action grounding
Action grounding requires an agent to tie proposed actions to current observable environment state, valid targets and verified preconditions instead of acting from stale assumptions or model-generated descriptions.
Mental model
Grounding closes the loop between perception and action. Before a consequential action, the agent should know what object or state it is acting on, how that state was observed, and whether the preconditions still hold.
Why it matters
Computer-use and tool-using agents can plan correctly yet act on the wrong element, account or stale page. Natural-language intent is not enough: execution needs identifiers, observations and post-action verification that connect the model's decision to the real environment.
01
Observe, bind, act and verify
Acquire fresh state close to the action, bind the intended target to stable identifiers or validated selectors, check required preconditions, then execute through a scoped capability. Afterward, observe the environment again and compare the expected state transition with what actually happened. Replan when the observation has changed instead of forcing the old action.
02
Example: an agent clicks the wrong 'Delete' button
A dashboard contains several rows with identical Delete labels. The model describes the correct customer in its reasoning but uses a positional selector after the table re-sorts. Grounding binds the operation to the customer's stable row ID, checks the displayed account name immediately before execution, and verifies that only the intended record disappears.
Common failure modes
- Executing from a screenshot or state observation that may already be stale.
- Using visual position or ambiguous labels when stable identifiers are available.
- Assuming a successful tool response proves the intended environment state changed.
Engineering heuristics
- Refresh high-consequence state immediately before acting.
- Bind actions to stable identifiers and explicit preconditions.
- Verify the post-action world, not only the tool call status.
Takeaways
- 01Correct intent does not guarantee a correctly grounded action.
- 02Grounding links model decisions to observable environment state.
- 03Action verification is part of the grounding loop.
Reading evidence
This records actions you actually took; it does not claim mastery, proficiency, or certification.
Used in
This Concept is reused across these canonical learning paths.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.