These examples run against the current 0.6.1 API. Versioned filenames record when a capability was introduced; they are not separate installation requirements. The roadmap describes upcoming enhancements.
From the repository root, install the development environment:
uv sync --extra dev --extra langgraph
All Python examples use local data or scripted responses and require no API keys.
The two LangGraph graph examples require the optional langgraph extra.
| Example | Demonstrates | Run |
|---|---|---|
| Customer support | Observe vs Govern, policy/risk/intent checks, BLOCK interception | uv run python examples/governed_customer_support_agent.py |
| LangGraph Observe | Real graph with a scripted model; advisory verdicts allow calls through | uv run python examples/langgraph_refund_observe_mode.py |
| LangGraph Govern | Real graph; $850 refund blocked, $120 refund allowed under a $500 constraint | uv run python examples/langgraph_intent_guardian_govern_mode.py |
| Evaluator API | Explicit plan ALLOW/REPLAN, heuristic CHALLENGE, mock Judge, pipeline composition | uv run python examples/v0_3_planner_critic_judge.py |
| Provider contract | Explicit offline MockJudge responses, swapping, validation, synthetic usage accounting | uv run python examples/v0_3_judge_providers.py |
| Budget and approval types | Manual usage, PAUSE verdict, caller-managed approval response, audit serialization | uv run python examples/v0_4_budget_and_escalation.py |
| Governance | Runtime approve-once, rejection, suspension, full-plan review and shared Judge budget | uv run python examples/v0_6_1_governance.py |
| Narration | Manually recorded decisions, compliance, budget values, and narrator pause/resume state | uv run python examples/v0_5_status_narration.py |
| CLI guide | Local status snapshots, follow formatting, JSON output, demo, and version | uv run agentic-sidecar --help |
| Interactive notebook | Deterministic constraints across support, operations, and research scenarios | Open with Jupyter after installing it separately |
evaluate() returns a decision; it does not execute or stop the action itself.
The LangGraph wrapper allows ALLOW/WARN in Govern mode; other outcomes resolve
or suspend. See governance walkthrough for runtime approvals.
Planner examples supply plan and goal directly to PlanEvaluator.evaluate().
Mock Judge responses and usage values illustrate the interface rather than
measured model quality, latency, or pricing. Narrator pause/resume updates its
reported state; it does not control an agent. Live CLI/runtime status is planned
for 0.7, and additional adapters/integrations for 0.8.
See the v0.8 scoped-agent, correlated-evidence and MCP gateway examples described in the examples guide.