AI Agent

The AI Agent is an interactive assistant built into Lifecycle that helps you debug and investigate issues in your ephemeral environments. It inspects Kubernetes resources, reads pod logs, queries the Lifecycle database, browses your GitHub repository, and can apply fixes directly.

Accessing the agent

Navigate to the AI Agent from any build’s detail view:

The chat interface is scoped to that build’s environment — the agent already knows which namespace, services, and resources to inspect.

Debugging a failing environment

Here’s a typical debugging session:

1. Open the AI Agent page for your build.
2. Click one of the suggested prompts — for example, “Why are my pods not starting?”
3. The agent queries Kubernetes for pod statuses, reads recent logs, and checks deployment configurations.
4. It returns a structured response with a summary, per-service findings, and suggested fixes.

At this point, you should see a response with service cards and an activity panel (described below). If the agent suggests a fix, you can apply it with one click.

The suggested prompts on first load are:

• “Why is my build failing?”
• “What’s wrong with deployments?”
• “Why are my pods not starting?”

You can also type your own question.

Understanding responses

The agent returns two response formats depending on the question:

• Plain text — for simple or conversational answers
• Structured investigation — for debugging queries, containing the sections below

Summary

A brief overview at the top describing overall environment health and key findings.

Service cards

Each service gets its own card with:

• Status chip — service health: healthy, degraded, failing, pending, or unknown
• Issue — root cause or current problem
• Errors — specific error messages from logs or events
• Suggested fixes — actionable steps you can follow
• Auto-fix button — applies a fix directly (for example, restarting a deployment or committing a config change to your PR branch)
• Evidence — links back to the tool calls that support the findings

Activity panel

Each response includes a collapsible Investigation panel showing every tool call the agent made:

• Status icon — spinner (in progress), green checkmark (success), or red X (failure)
• Description — what the agent did (e.g., “Fetched pods in namespace env-abc123”)
• Duration chip — how long the call took

A total investigation time appears at the top of the panel. Evidence references in the response link to specific entries in this panel.

Selecting a model

A Model dropdown in the chat header displays available models in the format provider:modelId (for example, “Claude Sonnet” or “GPT-4o”). Your selection is saved to localStorage and persists across sessions.

A Clear button appears when you have messages. It clears the conversation history and starts a new session.

Supported providers

Each provider requires its own API key configured on the server:

Agent capabilities

The agent has access to tools organized by category:

You can also extend the agent with external tools via MCP Integration.

Summary