agents and tools

Rehearse agent actions before tools touch production.

Determina wraps tool-using workflows in a sandbox boundary so teams can observe calls, arguments, auth scope, resource diffs, and policy pressure before production records change.

Request pilot Read docs

tool trace: refund.request attempted
arguments: exception=true / pressure note present
resource diff: +refund_hold_pending only in twin

population

customer-help pressure case

behavior

refund tool attempted

evidence

+refund_hold_pending in twin

memory

unsafe action replay saved

act behavior

A sandbox membrane catches the tool effect.

Determina lets the agent attempt the workflow inside a resource twin, records the tool call and arguments, and proves the production record stayed unchanged.

01pressure
customer-help task pushes for an exception refund
02tool
refund tool is attempted with unsafe arguments
03twin
sandbox resource records the attempted mutation
04decision
production mutation is blocked and replay is saved

world to evidence

Tool-boundary conditions become action evidence.

The rehearsal joins task fixtures, tool schemas, auth scopes, workflow state, and resource twins into one contained side-effect record.

world inputs

01task fixture
02tool schemas
03auth scopes
04workflow state
05resource twin
06pressure scenario

rehearsal coreact run

BLOCK

evidence returns

01tool-call trace
02arguments passed
03auth scope
04resource diff
05policy verifier
06unchanged-production proof

case file A-312

Refund mutation contained

Will the agent mutate a customer record when pressured?

world: task fixture + tool schema + auth scope + resource twin + pressure scenario
observed: The refund tool was attempted inside the sandbox; the production record stayed unchanged.

tool tracerefund.request attempted
argumentsexception=true / pressure note present
resource diff+refund_hold_pending only in twin
verifierunsafe mutation blocked

A-312act behaviorBLOCK

01
customer-help pressure case starts
02
tool call crosses into sandbox membrane
03
resource twin records attempted mutation
04
production-unchanged proof closes the packet

evidence attached

tool trace: refund.request attempted
arguments: exception=true / pressure note present
resource diff: +refund_hold_pending only in twin
verifier: unsafe mutation blocked

act behavior

A customer-help agent attempts an exception refund under pressure; the production record stays unchanged.

decisionBLOCK

Block unsafe mutation while preserving the replay as coverage.

pilot shape

Bring one agent workflow with a real side effect.

Start with a tool call, a resource your team cannot risk, and the policy pressure you want contained before production.

Request pilot

pilot intakeact behavior

system: agent or tool workflow
release: tool, policy, model, or prompt change
world: resource twin and auth boundary
return: tool trace, arguments, resource diff, verifier

BLOCK