Agents / demo
AGENTS.md1 starsChanged 58 days agoApache-2.0View on GitHub
A six-beat offline demo script. The numbers in it are spoken aloud, so they are load-bearing. A fully offline, deterministic, six-to-eight-minute runbook for a mixed engineer and leadership audience. README.md is the spoken script itself: what to type, what appears, what to say. This file is what an agent must know before changing anything the script narrates. - Beat 3 fails on purpose. configs/eval.fail.yaml differs from the passing config by one stricter threshold and must exit non-zero; run_demo.sh quarantines it…
In this repository39 other files it gives its agents.
- agent-core/AGENTS.mdAGENTS.md
- .agents/AGENTS.mdAGENTS.md
- AGENTS.mdAGENTS.md
- behavioral-regression/AGENTS.mdAGENTS.md
- claude-foundation/AGENTS.mdAGENTS.md
- config/AGENTS.mdAGENTS.md
- corpora/AGENTS.mdAGENTS.md
- docs/AGENTS.mdAGENTS.md
- examples/AGENTS.mdAGENTS.md
- experiments/AGENTS.mdAGENTS.md
- flow-corpus/AGENTS.mdAGENTS.md
- flow-protocol/AGENTS.mdAGENTS.md
- openspec/AGENTS.mdAGENTS.md
- scripts/AGENTS.mdAGENTS.md
- skills/AGENTS.mdAGENTS.md
- tests/AGENTS.mdAGENTS.md
- claude-foundation/CLAUDE.mdCLAUDE.md
- merge-gate-auditorSkill
- update-executive-reportSkill
- update-executive-reportSkill
- architecture-drift-guardSkill
- commonSkill
- corpus-guardianSkill
- dataset-lintSkill
- deploySkill
- e2e-matrix-sentinelSkill
- eval-corpus-forgeSkill
- hierarchical-recursive-brainstormSkill
- model-benchSkill
- openai-judgeSkill
- openspec-implementation-reviewSkill
- openspec-peer-reviewSkill
- openspec-quality-planSkill
- pre-pr-gateSkill
- project-setupSkill
- quality-gateSkill
- refactoring-decomposerSkill
- repo-invariant-reviewSkill
- test-completeness-guardSkill
The file
Files
AGENTS.md — demo
A six-beat offline demo script. The numbers in it are spoken aloud, so they are load-bearing.
A fully offline, deterministic, six-to-eight-minute runbook for a mixed engineer and
leadership audience. README.md is the spoken script itself: what to type,
what appears, what to say. This file is what an agent must know before changing anything
the script narrates.
Map
| Path | Role |
run_demo.sh | One-shot orchestrator for all six beats; writes to out/demo/ |
configs/eval.pass.yaml, configs/eval.fail.yaml | Same eval, one stricter threshold apart |
configs/compare.yaml, configs/live.appendix.yaml | Multi-model board; optional credentialed arm |
support_bot_target.py | The deterministic offline system under test |
data/support_bot.jsonl | Ten support questions |
replay/baseline.jsonl, replay_stubs.py | Recorded envelopes and the stale-search stub for beat 6 |
deck.html | Self-contained stakeholder walkthrough |
Diagram
flowchart LR
accTitle: the six demo beats and what each one proves
accDescr: The orchestrator runs list-plugins, a passing eval, a deliberately failing gate, a model comparison, a calibrated regression decision and a fixture replay, writing every report under out slash demo.
SH["run_demo.sh<br/>PYTHONPATH and allowlist"]
subgraph DM["demo/ (you are here)"]
B1["1 list-plugins"]
B2["2 eval.pass gate PASS"]
B3["3 eval.fail exit 1<br/>fails on purpose"]
B4["4 compare"]
B5["5 bregress ship hold escalate"]
B6["6 replay exact and counterfactual"]
end
OUT["out/demo/<br/>reports, gitignored"]
SH --> B1
SH --> B2
SH --> B3
SH --> B4
SH --> B5
SH --> B6
B2 --> OUT
B5 --> OUT
B6 --> OUT
classDef here fill:#e8f0fe,stroke:#1a73e8,stroke-width:2px
class DM hereRules that bite here
- Beat 3 fails on purpose.
configs/eval.fail.yamldiffers from the passing config by one stricter threshold and must exit non-zero;run_demo.shquarantines it so the strict shell mode does not abort. Do not "fix" it, and do not relax the threshold. - The quoted numbers are part of the artifact.
README.mdstates an n and a mean for the judged scorer. Change the target, the dataset or a scorer and those lines become false on stage. Re-run the beat and restamp the prose in the same change. - These configs are not under
../config/, so nothing forces a reviewer to see them. They are outside the protected path set by design, which puts the whole burden of keeping them offline and deterministic on whoever edits them. - The demo needs two environment variables and will not run without them.
PYTHONPATH=.so the callable target imports, andEVAL_HARNESS_CALLABLE_TARGET_ALLOWLIST=demobecause a config that names a module to import is executable input. Scoped todemo, never widened. - Never demo the auto-fix loop. It is deliberately disabled scaffolding (ADR 0004), and presenting it as working is the one claim this runbook cannot support.
Verify
bash demo/run_demo.shSubagents
| Task in this directory | Agent | Why |
| Find every place a quoted demo number also appears | explorer | Read-only; the same figure is repeated in the script, the deck and the root docs |
| Run the whole script and report which beat diverged | test-runner | Has Bash; the orchestrator is a shell script with six independent exits |
| Review a config or target edit before a live run | narrow-critic | A widened allowlist or a stray network call is invisible until it fails on stage |
See also
| Doc | Read it when |
README.md | You are running or rehearsing the demo and need the spoken script |
../docs/decisions/0004-auto-fix-loop.md | Someone asks why the auto-fix loop is not in the demo |
../docs/decisions/0049-fixture-replay-target.md | You are changing beat 6 and need why replay is a target and not a scorer |
../docs/e2e-live-journey.md | You need a live, non-mock generate-score-judge capture rather than this offline script |
../config/AGENTS.md | You are moving a demo config into the protected shipped set |
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

