Agents that talk to each other: Deep MCP Agent adds Cross-Agent Communication

🚀 Deep MCP Agent v0.5.0 — Cross-Agent Communication Arrives

Release Date: 2025-10-18

Deep MCP Agent 0.5 introduces Cross-Agent Communication, enabling one agent to call another as a tool — no extra servers, no orchestration layers, just pure MCP over HTTP/SSE.

✨ Highlights

• Cross-Agent Communication:
  Agents can now collaborate through new tools:

  • ask_agent_<name> – send a message to a specific peer and get its final answer.
  • broadcast_to_agents – query multiple peers in parallel and collect their replies.

• Plug-and-Play Peers:
  Attach any runnable agent (LangChain ReAct, LangGraph, or DeepAgents loop) with:

  cross_agents={"researcher": CrossAgent(agent=peer, description="Web research peer")}

• No Extra Infra: Runs entirely in-process; uses the same MCP tool ecosystem over HTTP/SSE.

🐛 Fixes & Improvements

• Clearer trace output for cross-agent tool calls.
• Better error messages for unknown peers in broadcasts.
• Sync _run path added to all tools to satisfy BaseTool requirements.
• Migrated to Pydantic v2 (ConfigDict) for clean, warning-free configuration.

📘 Docs & Examples

• New guide: Cross-Agent Communication — how to connect multiple agents seamlessly.
• Example: examples/use_cross_agent.py

⚙️ Compatibility

• Works with DeepAgents (preferred) or falls back to LangGraph ReAct prebuilt.
• Compatible with anyio >= 3.0 and Pydantic v2.
• Fully backward-compatible — if cross_agents is omitted, nothing changes.

GitHub: https://github.com/cryxnet/deepmcpagent
PyPI: pip install deepmcpagent

🧾 Changelog — v0.4.1 (2025-10-17)

🐛 Bug Fixes

• LangGraph compatibility:
  Fixed TypeError when falling back to create_react_agent() with langgraph>=0.6.
  The agent builder now dynamically detects supported parameters and omits deprecated ones, ensuring smooth operation across all LangGraph versions.
• Trace tool logging:
  Improved _after() logic to skip None values when trace_tools=True, correctly displaying tool results instead of None.

⚙️ Maintenance

• Updated runtime compatibility for environments without deepagents installed (graceful fallback to ReAct agent).
• Minor code cleanup and improved defensive checks in agent builder.

This release focuses on stability, developer experience, and documentation, making DeepMCPAgent easier to use in real projects. It includes bug fixes, tooling upgrades, CLI improvements, and better docs.

✨ Added

• CLI enhancements

  • deepmcpagent list-tools to inspect available MCP tools.
  • deepmcpagent run for interactive agent sessions.
  • Proper --version flag with dynamic version info from setuptools-scm.

• Improved tool tracing

  • Clear streaming trace of tool calls, arguments, and results in examples.
  • Human-readable final answers with intermediate steps preserved.

• Docs & Examples

  • Replaced mkdocs-click with a hand-written CLI reference (docs/cli.md).
  • Added working examples for both math and dictionary MCP servers.
  • Developer setup and contribution guide cleaned up and clarified.

🐛 Fixed

• Fixed missing langgraph.prebuilt imports by switching to langgraph-prebuilt.
• Fixed type errors across cli.py, agent.py, tools.py, and clients.py under strict mypy.
• Fixed Ruff violations (e.g., collections.abc.Callable, X | None annotations, contextlib.suppress).
• Fixed CLI option parsing (typer.Option default errors).
• Fixed broken MkDocs build pipeline by removing unused mkdocs-click plugin.

🛠️ Changed / Adapted

• Quality improvements across the whole project:

  • Enforced ruff + mypy strict as pre-commit standards.
  • Updated imports for Python 3.10+ style.
  • Hardened error handling in tool calls (fail-safe tracing, clear error messages).

• Project metadata

  • setuptools-scm for fully dynamic versioning (git tag → PyPI).
  • Updated classifiers and keywords.

• CI/CD pipeline

  • GitHub Actions workflow for lint, type-check, tests, build, and PyPI publish via Trusted Publishing.

📚 Docs & Contribution

• Clearer Contributing Guide with step-by-step dev setup, PR standards, and commit conventions.
• Added Changelog as part of the docs.
• MkDocs Material setup with clean navigation and working build.

Thanks to all contributors!

✅ Summary

DeepMCPAgent 0.4.0 is a stabilization + DX release:

• Easier to run agents and inspect tools.
• Clean, strict, and maintainable codebase.
• Docs and examples you can actually run out of the box.

DeepMCPAgent 0.3.0

🚀 Features

• CLI

  • Added global --version flag (prints package version).
  • Improved option parsing (--stdio / --http) with Typer Annotated.

• Tools

  • Improved JSON Schema → Pydantic conversion:

    • Carry through defaults and descriptions.
    • Generate per-tool arg models (Args_<tool>).
    • Sanitize model names for Pydantic compatibility.

🛠 Fixes

• Typing

  • Added return type annotations to CLI commands.
  • Parameterized Runnable and FastMCPClient to satisfy mypy.
  • Fixed create_model overload typing with safe casts.
  • Resolved Ruff B008 warnings (Typer option defaults).

• CI

  • Added PyPI Trusted Publishing workflow (publishes on tags).
  • Fixed workflow to skip deep extra on Python 3.10.
  • Migrated Ruff config to [tool.ruff.lint].
  • Applied ruff format so CI passes.

📦 Release

• Version bumped to 0.3.0.

• First automated release to PyPI via Trusted Publishing.
  Install with:

  pip install -U deepmcpagent