OpenWorker
#checkout-alerts@openworker checkout API is 500ing
Put AI to work.
Walk into your next meeting prepared.
#checkout-alerts›9 steps
Assistant
Likely regression in the 2:02 AM deploy
The error spike begins immediately after the payment migration change and matches the runbook’s rollback condition.
- Customer impact: checkout requests
- Rollback path is ready
▤
Checkout incident timelinecheckout-incident-timeline.md
Open ›
↗Send this report to #checkout-alerts?Leaves this Mac → Slack
ApproveNot now
←
···
checkout-incident-timeline.md
Checkout incident timeline
Payment migration · Updated 2:12 AM
18.4%peak error rate
10 mincustomer impact
Checkoutaffected service
Incident timeline
Deploy completed · payment migration v3.8
Alert fired · checkout 500s crossed 18%
Cause isolated · connection pool regression
Rollback verified · runbook checks passed
↩Recommended next actionRoll back v3.8, then monitor errors for 10 minutes.
Illustrative Desktop workflows
Build your own apps powered by AI agents.
Bundle with your app
Your application
Your UI & business logic
OpenWorkerHarness SDK + runtime
Ships as part of your app
Run agents alongside your product—in a desktop app or your server backend.
Deploy as a service
Your web app
Your backendHarness SDK
Secure API connection
OpenWorker Harness serviceAgent runtime
Run on Kubernetes in AWS, Azure or GCP
Connect your backend to a dedicated Harness service. Scale the runtime independently.
Keep agent actions aligned with the task using auto-approve.
Agent activity
You
Find out why the service is returning errors.
DevOps agent
I’ll check the logs and recent deployment.
I’ll continue investigating. The service was not restarted.
Action review by an LLM
Auto-approverChecks whether the proposed action aligns with the user’s task
You asked to
Investigate errors
Agent proposes to
Restart the service
Blocked before execution
The user only asked to investigate. Restarting the service makes a change they did not request.
The decision goes back to the agent so it can continue with the original task.
Select an activity to inspect its review. Auto-review uses a model; sandbox permissions are a separate control.
Give agent tools a workspace with boundaries.
Agent activity
You
Investigate the errors in this project.
DevOps agent
I’ll inspect the project and check its API.
I can continue with the project logs. Host credentials remain inaccessible.
Tool execution boundary
NVIDIA OpenShell sandbox
Tool process
read ~/.aws/credentialsAllowed by your policy
Project files
/workspace/projectStaging API
staging.example.comHost credentials
Outside sandbox~/.aws/credentialsAccess blocked
The host credential file is not mounted into this sandbox.
Select an activity to inspect its access. Applies to tools routed into OpenShell; your application configures filesystem and network permissions.
Keep sensitive content out of model input.
Agent activity
You
Read the application log and explain the error.
DevOps agent
I’ll read the log to find what failed.
- Tool result filtered1 secret redacted
- Check connection settingsDone
The request failed because the database did not respond within the configured timeout.
Tool output → model input
Tool output
read_filestatus=error reason=database_timeout api_key=ACME_demo_secret
Your guardrail1 match
Match secret pattern → replace with [REDACTED]
View rule
Tool
coding.read_filePattern/ACME_[A-Za-z0-9_]+/gActionredactDelivered to the model
status=error reason=database_timeout api_key=[REDACTED]
The error context stays useful. The matched secret is removed.
This example filters a tool result with a configured regex. Guardrails can also check user input and model output; detection depends on the rules you configure.
Your app. Your choice of models.
OpenAI
Anthropic
Gemini
xAI
Mistral
DeepSeek
Kimi
MiniMax
Qwen
OpenRouter
Together AI
Fireworks
Ollama
vLLMUse a provider directly or through a compatible gateway.
Keep your app experience the same.