Examples

Eight agents under examples/ run end to end with no SaaS account and no model key. Each has a README and a test (make test-e2e-examples; the data analyst has its own target) that runs it against recorded GitHub, Sentry, PagerDuty, Calendar and Slack servers and a fixture model, plus the real surface gateway for chat. Only the Telegram and Twilio paths need real accounts.

Example	Shows	Path
Support desk	Answers from your docs, hands off to a human	examples/support-desk/
PR triage	A webhook trigger, GitHub and Slack actions	examples/pr-triage-webhook/
Incident responder	Two triggers and an approval node	examples/incident-responder/
Daily digest	A timezone-aware schedule and a hard budget	examples/daily-digest/
Telegram assistant	The surface gateway and a calendar action	examples/telegram-assistant/
Voice receptionist	Twilio under a hard spend cap	examples/voice-receptionist/
Data analyst	pandas in a microVM, deny-by-default egress	examples/data-analyst/
Eval-gated deploy	A CI gate that blocks a regression	examples/eval-gated-deploy/

The REST examples assume export API=http://localhost:8080 TOKEN=<JWT or API key> and send -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json". Start the stack first (Develop locally).

Support desk

A webchat agent answers from your docs, shows its source, and hands the conversation to a human when it cannot answer, pinging Slack. The APIs are in Your documents and memory.

make dev-infra && make run-api          # terminal 1
python3 examples/support-desk/setup.py   # creates the agent, uploads the docs
# create an API key with scope runs:execute (Settings, API keys), then open examples/support-desk/demo.html

Ask How do I reset my password? for a cited answer, then I want a refund: the widget shows “Connecting you to a human”. An operator replies with POST /v1/sessions/{id}/operator-messages and the reply appears in the widget. setup.py deploys over REST because it also uploads the documents and mints the widget key.

PR triage to a Slack thread

Every opened GitHub pull request is read, triaged by a model in one line, commented on the PR, and announced in Slack with the verdict in the thread. It uses a webhook trigger, the GitHub get_pr and comment_pr actions and Slack post_message and reply_thread, in a workflow.json graph.

cd examples/pr-triage-webhook
curl -X POST $API/v1/agents -H "Authorization: Bearer $TOKEN" -d '{"name":"pr-triage-webhook","description":"PR triage","systemPrompt":"Workflow agent: see workflow.json"}'
curl -X PUT $API/v1/agents/pr-triage-webhook/workflow -H "Authorization: Bearer $TOKEN" -d @workflow.json
curl -X POST $API/v1/agents/pr-triage-webhook/triggers -H "Authorization: Bearer $TOKEN" -d @trigger.json
# -> {"id": "<trigger-id>", "secret": "<shown once>"}
./send-fixture.sh $API/v1/hooks/<trigger-id> <secret>   # signed pull_request.opened

For real GitHub, install the github and slack connectors, then add a repository webhook with payload URL $API/v1/hooks/<trigger-id>, content type JSON, the trigger secret and the Pull requests event. A re-sent delivery returns 200 duplicate and starts no second run. A bad signature is 401; an agent over its budget gets 402.

Incident responder with an approval

A Sentry alert or PagerDuty incident is summarized with the Sentry issue and PagerDuty incident in hand, a runbook step is posted to Slack, and the run parks at an approval node. Approving posts a note on the PagerDuty incident. Denying fails the step and posts nothing. See Holds and trust for the approval mechanism.

cd examples/incident-responder
# create the agent and PUT workflow.json as above, then create both triggers
curl -X POST $API/v1/agents/incident-responder/triggers -H "Authorization: Bearer $TOKEN" -d @trigger-sentry.json    # and trigger-pagerduty.json
./send-fixture.sh sentry $API/v1/hooks/<sentry-trigger-id> <secret>
# the run parks; approve or deny it:
curl -X POST $API/v1/runs/<run-id>/takeover/<takeover-id>/grant -H "Authorization: Bearer $TOKEN" -d '{"notes":"go"}'
curl -X POST $API/v1/runs/<run-id>/takeover/<takeover-id>/deny -H "Authorization: Bearer $TOKEN" -d '{"notes":"not now"}'

Daily digest to Slack

Every weekday at 09:00 New York time, the agent lists open GitHub PRs, has a model write a short digest and posts it to Slack. It is the smallest unattended agent: a schedule with a timezone, a three-node workflow and a hard budget cap.

cd examples/daily-digest
pal deploy                      # agent + system prompt + workflow from agent.yaml
# install the slack and github connectors, edit owner/repo/channel in schedule.json, then:
curl -X POST $API/v1/schedules -H "Authorization: Bearer $TOKEN" -d @schedule.json   # nextFireAt = next weekday 09:00 NY

pal deploy covers the agent. The schedule has no deploy field, so you create it over REST. Run it now with POST /v1/runs. The scheduler polls every 60 s.

Telegram assistant

Message a Telegram bot (or WhatsApp Cloud number) and a workflow reads your Google Calendar and answers in the same chat. It uses the surface gateway, the Telegram secret-token and WhatsApp X-Hub-Signature-256 checks (Set up PAL) and the calendar list_events action.

cd examples/telegram-assistant
pal deploy
curl -X POST $API/v1/runs -H "Authorization: Bearer $TOKEN" -d '{"agentName":"telegram-assistant","input":{"text":"hi"}}'   # priming run
# run the gateway (see the README), then setWebhook with secret_token, and message the bot

The priming run is needed because the gateway starts runs over gRPC, which refuses an agent that has never had a REST run (no promoted version). Replies are plain text, once, within 3 minutes.

Voice receptionist

A Twilio number answers inbound calls under a hard spend cap. Forged webhooks get 403. An over-budget call is declined with <Reject/> before it connects. The status callback reconciles the $0.05 reservation to the real per-started-minute cost. Details are in Set up PAL.

cd examples/voice-receptionist
pal deploy
curl -X PUT $API/v1/agents/voice-receptionist/budget -H "Authorization: Bearer $TOKEN" -d '{"maxCostUsdPerDay":2,"hardFail":true}'
curl -X POST $API/v1/voice/numbers -H "Authorization: Bearer $TOKEN" -d @number.json    # your SID, auth token, number

Budgets and voice numbers are not part of the deploy manifest, so they stay on REST. The inbound reply is a fixed greeting; the realtime conversation needs the LiveKit media worker.

Data analyst in a sandbox

Upload a CSV, run pandas in an isolated microVM and get JSON plus a chart back; an agent then explains the result. The code that touches the data never runs in the control plane, and egress is deny-by-default. See Sandbox lifecycle.

docker build -t pal/examples/data-analyst examples/data-analyst   # then use the image digest below
VM=$(curl -s -XPOST $API/v1/runtime/schedule -H "Authorization: Bearer $TOKEN" \
  -d '{"imageDigest":"<sha256:...>","isolation":"untrusted","network":"allowlist_domain"}' | jq -r .vmId)
curl -XPUT "$API/v1/runtime/vms/$VM/files/content?path=sales.csv" -H "Authorization: Bearer $TOKEN" --data-binary @sales.csv
curl -XPOST $API/v1/runtime/vms/$VM/exec -H "Authorization: Bearer $TOKEN" \
  -d '{"command":"python","argv":["/opt/analyze.py","/workspace/sales.csv"]}'

make test-e2e-analyst runs TestDataAnalystSandbox against a stack with the scheduler and runtime manager running; it is not part of make test-e2e-examples. The default tier uses the manager’s docker backend with no network and a fixture model. The files API is Firecracker-only, so on the default tier the CSV goes in through exec. The Firecracker tier (needs /dev/kvm) adds the files-API upload and an audited egress deny.

Eval-gated deploy

A CI pipeline that cannot ship a worse agent. It runs an eval suite against the candidate and records it with POST /v1/eval-runs. A 422 (regressed against the pinned baseline) stops the job before pal deploy --json. A passing change is deployed, a receipt is issued with POST /v1/runs/{id}/receipt and verified at POST /v1/runs/receipts/verify. The folder holds scripts and a sample GitHub Actions workflow. See Evaluations.

cd examples/eval-gated-deploy
EVAL_PIN=1 scripts/eval.sh .                    # baseline: 3/3 cases, pinned for branch main
scripts/eval.sh candidates/regressed; echo $?   # 1: HTTP 422, nothing deploys
scripts/eval.sh candidates/improved && scripts/release.sh candidates/improved

The server stores the per-case results you post and compares the mean score. The script scores each case (assert.contains) from the run output.

Limits

• A model reply is dropped into JSON-free connector parameters, so the prompts ask for one plain line. Keep that if you edit them.
• The incident trigger templates pin the Sentry organization and the PagerDuty incident or issue id to fixture values. Edit trigger-*.json for your account.
• PR comments use the PR conversation (comment_pr). Inline review comments on a diff line are not wired.