AgentGuard Proof Gallery

These are local, runnable proof paths for the public SDK. Use them in demos, issues, discussions, launch posts, and first-run validation.

No hosted dashboard or API key is required unless a command explicitly says so.

1. Offline Guard Demo

pip install agentguard47
agentguard demo

Proves:

• BudgetGuard stops simulated spend
• LoopGuard stops repeated tool calls
• RetryGuard stops retry storms
• no network calls are required

Expected stop condition:

Local proof complete.

2. Coding-Agent Review Loop

python examples/coding_agent_review_loop.py
agentguard incident coding_agent_review_loop_traces.jsonl

Proves:

• repeated review/edit attempts can burn budget
• stuck patch retries are stopped locally
• the same trace can become an incident report

Sample incident: docs/examples/coding-agent-review-loop-incident.md

3. Sticky Agent Proof

python examples/sticky_agent_proof.py --out-dir proof/sticky-agent-proof
agentguard incident proof/sticky-agent-proof/sticky_agent_proof_traces.jsonl

Proves:

• one CrewAI-style workflow can show retry storm, loop detection, and budget burn
• the SDK can write a local incident report and hosted-compatible NDJSON
• dashboard proof can be validated without adding framework dependencies

Expected artifacts:

sticky_agent_proof_traces.jsonl
sticky_agent_proof_hosted.ndjson
sticky_agent_proof_incident.md

4. Token-Budget Spike

python examples/per_token_budget_spike.py
agentguard report per_token_budget_spike_traces.jsonl

Proves:

• one oversized context or completion can exceed a run budget
• the SDK can price a local synthetic turn without provider credentials
• spend control is not only about loop count

Expected stop condition:

Cost budget exceeded:

5. Decision Trace Workflow

python examples/decision_trace_workflow.py
agentguard decisions decision_trace_workflow.jsonl

Proves:

• agent proposals can be captured as decision.proposed
• human edits and overrides can carry reason/comment fields
• approval and binding outcomes use the normal trace pipeline

Expected event types:

decision.proposed
decision.edited
decision.approved
decision.bound

6. Budget-Aware Escalation

python examples/budget_aware_escalation.py

Proves:

• apps can keep a cheaper default model
• hard turns can be marked for escalation by deterministic rules
• the SDK advises without hiding provider routing inside AgentGuard

Expected stop condition:

Escalation reason: token_count 2430 exceeded 2000

7. Starter File Smoke Test

agentguard quickstart --framework raw --write
python agentguard_raw_quickstart.py
agentguard report .agentguard/traces.jsonl

Proves:

• a new repo can get a runnable local integration in one command
• generated starters write traces to .agentguard/traces.jsonl
• first value does not require a dashboard account

7. MCP Read Path

npm --prefix mcp-server test
AGENTGUARD_API_KEY=ag_... npx -y @agentguard47/mcp-server

Proves:

• the MCP server is a narrow read-only surface over retained AgentGuard data
• coding agents can inspect traces, decisions, alerts, usage, costs, and budget health after hosted ingest is enabled
• local SDK enforcement remains independent of the hosted dashboard

Expected boundary:

Requires AGENTGUARD_API_KEY for retained hosted data. Does not add local runtime enforcement.

What To Share

The most shareable public demo for the release train is the sticky agent proof:

python examples/sticky_agent_proof.py --out-dir proof/sticky-agent-proof
agentguard incident proof/sticky-agent-proof/sticky_agent_proof_traces.jsonl

It maps directly to the product wedge: a production-style agent workflow retries, loops, burns budget, then AgentGuard produces a local incident and a dashboard-ready event stream.