AI Agent
How the Strayl AI agent works — models, streaming reasoning, and context.
The Strayl AI agent is the core of the platform. It reads your codebase, interprets your instructions, writes code, runs commands, and iterates based on results — all in a continuous loop until the task is complete.
How the agent works
The agent follows a ReAct loop:
- Reason — think about the current task and what needs to be done
- Act — call a tool (write a file, run a command, search the codebase)
- Observe — read the tool output
- Repeat — continue until the task is complete or human input is needed
This loop runs transparently in the Chat panel. You can watch the agent's reasoning stream in real time.
Streaming reasoning
Strayl streams the agent's thinking as it works. You'll see:
- The agent's step-by-step reasoning before each action
- Tool calls with their inputs and outputs
- A final summary of what was done
This transparency lets you catch mistakes early and course-correct with a follow-up message.
Context window
The agent maintains context across your entire session. It knows:
- The full directory structure of your project
- The content of files it has read
- The complete conversation history
- Results of previous commands
For large codebases, the agent uses smart context retrieval — searching for relevant files rather than loading everything at once.
Available tools
The agent has access to the following tools in the project workspace:
| Tool | Description |
|---|---|
ls | List directory contents |
read | Read a file |
write | Write or create a file |
edit | Make targeted edits to a file |
grep | Search file contents |
glob | Find files by pattern |
exec | Run any shell command |
getLogs | Read dev server or process logs |
showPreview | Open the live preview in the browser |
deleteFile | Delete a file or directory |
moveFile | Move or rename a file |
Model selection
You can switch the underlying model from the chat toolbar. Strayl supports multiple frontier models. The default is optimized for coding tasks.
Tips for getting the best results:
- Be specific about what you want ("Add a
POST /api/usersendpoint that validates email with Zod") rather than vague ("Add an API") - If the agent goes in the wrong direction, say "Stop. Revert that and instead..." — the agent understands corrections
- For large refactors, use Plan Mode first to review the approach
- Paste error messages directly — the agent reads stack traces and fixes them
Interrupting the agent
If the agent is doing something wrong, type a new message to interrupt it. The current tool call will complete, then the agent will read your correction and adjust.