Field guide · shareable

AI agent safety checklist before production

Agents fail loudly (loops) or quietly (spend). Use this list whether you run Tollgate or something else. The point is evidence, not vibes.

Protect — hard limits

☐ Per-consumer budget (daily/monthly) with a hard deny when empty
☐ Tool-loop ceiling (max_tool_calls / estimated tool rounds) before the model starts
☐ Freeze / kill switch that stops admission without redeploying code
☐ Scopes — agents only get tools they need
☐ Secrets never live in agent prompts; keys stay off the wire to workers when possible

Route — survive outages

☐ At least two providers (or local + cloud) for critical paths
☐ Health-aware failover, not “retry the same dead host forever”
☐ Circuit breakers with cool-down (and optional jitter)
☐ Clear error to the desk when nothing is healthy (fail closed)

Prove — evidence

☐ Chaos / DR drill that kills a provider and checks recovery
☐ Scorecard you can show (resilience / certificate), not a slide
☐ Audit trail: who spent, who was denied, who froze
☐ 10-minute cold path a stranger can run without a sales call

Desk hygiene (Gnom-Hub and friends)

☐ Brainstorm chat ≠ auto-spend workers
☐ Explicit Execute before tools/files cost money
☐ Dry-run computer-use until a deliberate God-Mode

Run the Tollgate cold path (10 min)

Protect demo needs no API keys.

cold path

git clone https://github.com/landjunge/tollgate.git && cd tollgate
python3 -m venv .venv && .venv/bin/pip install -e .
./scripts/ten-minute.sh
# Dashboard: http://127.0.0.1:8787/dashboard

Share this list

Permalink: landjunge.github.io/tollgate/blog/checklist.html
Stack: Launch post · Gnom-Hub · Press kit

Ran the cold path? Drop your resilience score in Discussions — that is the best free marketing we have.