Demo history · DataPilot and ReturnPilot runs marked “simulated” come from af-run demo — simulated agent behaviour scored by the real graders, statistics and gate; not measurements of the live apps.

System · 11

Settings

Keys, budgets, the agents registry and the health of every connection. Admin actions are confirmed and audit-logged.

Connections

what this deployment can reach

Neon Postgres
DATABASE_URL set · 16.9 MB of 1 GB free plan
GitHub sign-in (Auth.js)
AUTH_GITHUB_ID / AUTH_GITHUB_SECRET set
GitHub token (dispatch · statuses · PR comments)
GH_TOKEN valid for the AgentForge repo
Runner key
RUNNER_KEY_HASH set — GitHub Actions can post results
Nightly schedule
last nightly 10 days ago — over 36 h: GitHub disables schedules after 60 days without activity

database size17 / 1,024 · 2%

Daily LLM budgets

ledger · 2026-10-09

Rubric judge · Gemini Flash9 / 150 · 6%

Cheap graders · gpt-oss-20b0 / 400 · 0%

Prompt Guard 2 scoring0 / 500 · 0%

Mutation · reflection · drafts0 / 60 · 0%

Embeddings · mining0 / 500 · 0%

Target-agent eval calls

API keys

afk_live_… · SHA-256 stored · shown once · scoped to one agent

Name	Agent	Prefix	Scopes	Last used	Status
datapilot PR gate	DataPilot	afk_live_a2db…	gate:trigger	13 days ago	active
returnpilot PR gate	ReturnPilot	afk_live_d3c8…	gate:trigger	16 days ago	active
returnpilot live traces	ReturnPilot	afk_live_1755…	traces:writeprofiles:read

Agents registry

targets under test · eval adapter contract §S.4

DataPilot

gemini-3-flash-preview

repo: Sohail-5678/Datapilot-Multi-Agent-Data-Analyst
adapter: python -m datapilot.eval_adapter run …
workdir: backend
locked: sql_guard, read_only, scan_confirm_rows, sandbox_limits, output_grounding, budget_ceilings, pii_policy

ReturnPilot

gemini-3-flash-preview

repo: Sohail-5678/returnpilot
adapter: python -m returnpilot.eval_adapter run …
workdir: backend

Connections

Daily LLM budgets

API keys

Agents registry

Audit log