Deployment Agent
The Deployment Agent is a read-only AI inspector for your live deployments. Open it from the header Agent button.
It answers questions by fetching your deployment data — status, history, health — through tools, and can pull platform docs (via Moss search) when troubleshooting guidance is needed. It does not guess repos, services, or health states.
What it can do
| Capability | Example questions |
|---|---|
| List deployments | "Show me my deployments" |
| Inspect current state | "What's the status of my api service?" |
| Review history | "Why did my last deployment fail?" |
| Check health | "Is my service healthy right now?" |
| Search platform docs | "What does an ALB unhealthy target mean?" (often paired with a deployment tool) |
| Search external docs | "How does the ECS deployment circuit breaker work?" |
What it cannot do
The agent is read-only. It cannot:
- Trigger deploys or rollbacks
- Change env vars, branch, or subdomain
- Re-run Smart Analysis or Improve scan
- Access another user's deployments
For actions, use the deploy workspace UI.
How to ask effective questions
Use repo and service names from your dashboard:
| You say | Agent uses |
|---|---|
acme/smart-deploy | repoName: smart-deploy |
service web | serviceName: web |
Good prompts:
- "Show me my deployments"
- "Why did smart-deploy web fail on the last deploy?"
- "Is shop-api healthy right now?"
Ambiguous prompts: If you mention only api without a repo, the agent lists deployments first instead of guessing.
Tools the agent uses
| Tool | Returns |
|---|---|
list_deployments | Up to 25 deployments: status, branch, target, URL |
get_deployment_details | Status, commit, revision, region, cloud resources, scan summary |
get_deployment_history | Recent attempts: success/fail, failed step, log excerpts |
get_runtime_health | Recent probes: app status, HTTP code, latency, ECS/ALB signals |
search_docs | Relevant Smart Deploy doc excerpts (Moss + deterministic search) for troubleshooting and how-to |
search_external_docs | One public external technical-doc result, searched and converted to Markdown by Context.dev |
External documentation setup
Set CONTEXT_DEV_API_KEY on the WebSocket worker environment. The key remains server-side: the agent sends it only to Context.dev and never to the browser, model prompt, tool result, or logs. The tool is optional; without the key it returns a clear unavailable result and the rest of the agent continues to work.
External results are limited to one HTTPS page per tool call, with article content trimmed before it is sent to the model. Prefer official sources such as docs.aws.amazon.com; the tool can restrict a query to a supplied public hostname.
Live status updates
While working, the agent streams progress:
- Accepted → status updates → tool started/completed → final message
If the WebSocket worker is offline, you see: "The deployment agent is offline right now. Refresh the page and try again."
Limits
| Limit | Value |
|---|---|
| Tool calls per question | 6 |
| Conversation memory | Last 6 turns |
| History/health samples per tool | 5 |
| Write actions | None |
Complex root-cause analysis may hit the tool limit. Use Analyze failure on a specific history entry or read full logs in the History tab.
Starter prompts
Built-in shortcuts in the agent sheet:
- Show me my deployments
- Why did my last deployment fail?
- Is my service healthy right now?
When to escalate
| Need | Use instead |
|---|---|
| Full step logs | Deployment History tab |
| Deep failure analysis | Analyze failure on a history entry |
| Fix build plan | Improve scan |
| Platform how-to | FAQ, Debugging Deployments |
See AI Assistance.