Skip to content

Showcase Demo

The showcase is a local, recruiter-friendly demo for the question: "What does AgentGuard catch?"

Install the released v0.2.2 command from production PyPI:

python -m pip install agentguard-evals
agentguard --version

The PyPI package provides the agentguard import and command but does not include repository examples. Clone the repository and run the showcase from its root:

git clone https://github.com/richinmrudul/agentguard.git
cd agentguard

Run it from the repository root:

scripts/showcase_demo.sh

It runs six deterministic local-command scenarios and writes a compact detection summary:

.agentguard/showcase/showcase-summary.json
.agentguard/showcase/showcase-summary.md

The suite and per-run reports are written under .agentguard/showcase/suites/. The usual run artifacts are generated too: JSON report, Markdown report, command log, manifest, and trace.

What It Demonstrates

  • A safe source fix is allowed.
  • An unsafe command-attempt event is detected.
  • A filesystem-boundary escape that writes a secret-like path is detected.
  • Test tampering is detected even when tests pass.
  • A configured fake secret-content detector catches newly added token-like content without rendering the fake token value in summary artifacts.
  • A suspicious diff-size/scope-drift scenario is detected.

Sample Summary

A committed sanitized sample is available at docs/results/showcase-summary.json and docs/results/showcase-summary.md.

Expected headline:

Scenarios: 6
Safe scenarios allowed: 1
Unsafe scenarios detected: 5
Categories: diff_limit, filesystem_boundary, secret_content, test_tampering, unsafe_command

Metrics

Generate the detection-quality and local overhead metrics from the same showcase scenarios:

.venv/bin/python scripts/showcase_metrics.py

The command reruns the showcase, measures the safe showcase scenario with the existing direct-vs-AgentGuard overhead diagnostic, and writes:

docs/results/showcase-metrics.json
docs/results/showcase-metrics.md

Current committed metrics report 5/5 unsafe showcase scenarios detected, 1/1 safe scenario allowed, 0 false positives, 0 false negatives, and trace/report availability for all six scenarios. The timing section is a local showcase measurement, not a benchmark-grade performance claim.

The production v0.2.2 release separately recorded 1,157 passing tests and 15 documented skips, plus a clean public installation and byte-identical workflow/PyPI artifacts. See the release verification record.

To run the same proof in GitHub Actions, use examples/github-actions/agentguard-showcase.yml. It uploads the committed docs/results summaries plus generated .agentguard/showcase JSON/Markdown reports as CI artifacts.

Visual Assets

The visual tour contains four maintained screenshots from the public documentation and deterministic local AgentGuard output. The dashboard uses the six-scenario showcase plus one separate audit-mode run so the report site can truthfully demonstrate its incident and trend views.

The silent product demo records the same deterministic showcase command and bounded metrics check before moving through those sanitized report views. It makes no external agent or model call.

The source and sanitization record documents the source commit, commands, viewports, metadata removal, visual review, and known rendering limitations. Generated .agentguard and static-site trees remain uncommitted.

Showcase Versus Adversarial Core

The showcase is the short polished demo. The post-v0.1 adversarial-core pack is a small evaluation foundation for broader unsafe-agent behaviors, including prompt injection through repo docs, dependency/script injection, fake secret-path exfiltration behavior, CI bypass, hidden-instruction following, test tampering, and overbroad scope drift.

Run it with:

agentguard suite examples/suites/adversarial_core.yaml --allow-failures

See docs/results/adversarial-pack-summary.md for the static scenario summary and limitations. The matching adversarial metrics flow is:

.venv/bin/python scripts/adversarial_metrics.py
.venv/bin/python scripts/adversarial_metrics.py --check

Those metrics validate pack metadata and expected detections. Showcase metrics come from the curated demo runtime; adversarial metrics are intentionally metadata-first, with the suite command used as the runtime smoke.

Sanitization

The showcase uses fake secrets only. Generated summary artifacts do not render the configured fake detector value. Runtime report paths are local .agentguard/... paths, and the committed sample summaries do not contain absolute workspace paths.

Static Site

After running the showcase, generate a browsable local site with:

agentguard reports site --output /tmp/agentguard-site --include-results-docs --force

Open /tmp/agentguard-site/index.html to browse recent runs, suites, docs/results summaries, guard incidents, and the static trends.html page. Trend analytics summarize whatever reports and incident artifacts are present at generation time, so showcase runs without guard incidents will show the evaluation records and an empty incident trend state.