Skip to content

Agent Exploration API

The exploration API is the internal capability surface that lets agents behave like first-class workspace operators without unrestricted write access.

Rule

Agents should have:

  • broad read/explore permissions
  • constrained draft and execution permissions
  • explicit human approval for risky actions

Core functions

FunctionPurposePermissionApprovalLogging
list_workspace_objectsEnumerate objects by type and scopeReadNosummary
read_objectRead one object in fullReadNosummary
search_objectsSearch workspace objectsReadNosummary
suggest_workflowPropose a workflow designSuggestNofull
create_draft_lensCreate a draft lensDraftNofull
create_draft_agentCreate a draft agentDraftNofull
create_draft_toolCreate a draft toolDraftNofull
create_draft_workflowCreate a draft workflowDraftNofull
simulate_runDry-run an object locallySimulateNofull
run_evaluationExecute an evaluation suiteExecuteMaybefull
prepare_private_battleDraft a private battle specDraftNofull
generate_reportProduce a report artifactExecuteNofull
request_human_approvalEscalate a risky actionSuggestNofull

Safety stance

Agents should be able to:

  • inspect objects freely
  • draft new objects
  • propose changes
  • simulate work

Agents should not silently:

  • publish publicly
  • destroy data
  • send external messages
  • rotate provider credentials
  • exceed workspace cost rules

Audit requirements

Every non-read action should log:

  • actor type and actor id
  • workspace id
  • action verb
  • target object
  • approval ticket id if present
  • cost and latency if applicable
  • result status
  • version references