System Types

Agent System Type

Review agent trajectories for tool misuse, grounding gaps, refusal failure, state loss, unsafe effects, and external calls.

Risks Determina tests

Agent Behavior Trials review trajectories where final-output checks can miss the failure:

tool misuse
grounding gaps
refusal failure
multi-turn state loss
unsafe effects
external call pressure
latency cliffs

Supported drivers

Supported agent setup lanes include default HTTP session services, custom HTTP-session driver configs, in-process Python or LangGraph-style objects, OpenAI-compatible Chat Completions, Anthropic Messages, and MCP stdio.

shell

determina audit --project-id <project-id> --system-version-id <system-version-id> --scenario current-info-tool-use --output-dir ./agent-audit

Local and hosted boundaries

Agent docs must be explicit about lane and maturity. Local task/environment packs, generated agent packs, stress-swarm planning, external call variants, and corpus review are advanced/gated workflows, not default installed-package execution paths. They do not imply generated customer agents, hosted sandboxed side effects, managed production effectors, hidden tool interception, or automatic release approval.