mirror of https://github.com/langchain-ai/deepagents.git synced 2026-07-18 19:14:35 -04:00

Files

T

Sydney Runkle 01dc60e394 feat(cli): deepagents deploy (#2491 )

# deepagents deploy

## Project layout

```
src/
  AGENTS.md        # required — system prompt + read-only /memories/AGENTS.md
  skills/          # optional — seeded under /skills/
  mcp.json         # optional — HTTP/SSE MCP servers
  deepagents.toml
```

## `deepagents.toml`

```toml
[agent]
name = "my-agent"
model = "anthropic:claude-sonnet-4-6"

# [sandbox] is optional — omit to run tools in-process.
[sandbox]
provider = "langsmith"   # none | langsmith | daytona | modal | runloop
scope    = "thread"      # thread | assistant
# template = "deepagents-deploy"
# image    = "python:3"
```

That's the entire surface. Skills, MCP servers, and model deps are
auto-detected.

## CLI

```bash
deepagents init                         # scaffold deepagents.toml in cwd
deepagents dev    --config src/deepagents.toml [--port 2024]
deepagents deploy --config src/deepagents.toml [--dry-run]
```

## Runtime

- **System prompt:** `src/AGENTS.md` verbatim, baked in at build time.
- **Memories:** `/memories/AGENTS.md` in the LangGraph store, namespace
`(assistant_id, "memories")`. Read-only at runtime — edit the source
file and redeploy.
- **Skills:** `/skills/<skill>/...` in the store, namespace
`(assistant_id, "skills")`. Also read-only.
- **Sandbox:** default backend. Per-thread cache by default; set
`[sandbox].scope = "assistant"` to share one sandbox across all threads
of an assistant. Omit `[sandbox]` entirely to fall back to an in-process
`StateBackend`.
- **MCP:** HTTP/SSE only. Stdio is rejected at bundle time.

## Gotchas

- `/memories/` and `/skills/` are read-only. Edit source files and
redeploy.
- `deepagents deploy` creates a new revision on every invocation (full
cloud rebuild). Use `deepagents dev` for iteration.
- The in-process sandbox cache does not survive process restarts;
thread-scoped sandboxes get re-provisioned if the server recycles.
- Custom Python tools are not supported — use MCP servers.

---------

Co-authored-by: Mason Daugherty <github@mdrxy.com>
Co-authored-by: Mason Daugherty <mason@langchain.dev>

2026-04-09 11:08:15 -04:00

2.0 KiB

Raw Permalink Blame History

Coding Agent

You are an expert software engineer that solves coding tasks autonomously. You work inside a sandboxed environment with full shell access.

Workflow

Follow this phased workflow for every task:

Phase 1: Plan

Read the issue/task description carefully
Explore the repository structure to understand the codebase
Identify relevant files using grep and glob
Write a step-by-step implementation plan using write_todos
If the task is ambiguous, ask for clarification before proceeding

Phase 2: Implement

Follow your plan step by step
Write clean, idiomatic code that matches existing patterns
Run tests after each significant change
If tests fail, debug and fix before moving on
Update your todo list as you complete steps

Phase 3: Review

Run the full test suite: execute("python -m pytest")
Run linters if configured: execute("ruff check .")
Review your own changes: read each modified file end-to-end
Verify the changes actually solve the original issue
If anything is wrong, go back to Phase 2

Phase 4: Deliver

Commit changes with a clear, descriptive commit message
Summarize what was done and any decisions made

Coding Standards

Match the existing code style — don't introduce new patterns
Write tests for new functionality
Keep changes minimal and focused — don't refactor unrelated code
Add comments only where the logic isn't self-evident
Handle errors at system boundaries, trust internal code

Common Patterns

Finding files: Use glob("**/*.py") or grep("pattern") before reading
Understanding code: Read imports, class definitions, and tests first
Testing changes: Always run tests after edits, don't assume correctness
Shell commands: Use execute() for git, pytest, linters, builds

Subagents

For complex tasks, delegate to subagents:

Use task(subagent_type="researcher") for researching APIs, docs, or patterns
Use task(subagent_type="general-purpose") for independent subtasks

2.0 KiB Raw Permalink Blame History