Each AI framework package has isolated dependencies and its own virtual environment. This allows testing frameworks with incompatible library versions side-by-side.
For deep dives into individual clients, execution flows, and telemetry schemas, see:
- packages/IMPL.md — Implementation breakdown of each client.
- packages/FLOW.md — Control loop and logical boundary comparisons.
- packages/shared/OBSERVABILITY.md — Hook APIs and standardized token event schemas.
What each framework provides out-of-the-box, and how we use it:
| Capability | Pydantic AI | Agno | Strands | ADK |
|---|---|---|---|---|
| Loop driver | agent.iter() autonomous node loop |
Explicit async while — one agent.arun() per turn |
agent(prompt) single synchronous call |
runner.run_async() event stream |
| MCP integration | MCPToolset (native) |
MCPTools (native) |
MCPClient (native, multi-transport) |
McpToolset (native, tool_filter support) |
| Observability hooks | Usage delta from agent_run.usage each node |
post_hooks=[fn] on Agent |
6 typed lifecycle events (BeforeInvocationEvent … AfterToolCallEvent) |
after_model_callback on Agent |
| Model routing | KnownModelName strings (google-gla:, anthropic:, …) |
Native model objects (Claude, Gemini, OpenAIChat, Ollama) |
LiteLLM model strings — any provider | Native Gemini ID or LiteLlm(model=…) |
| Context/budget control | UsageLimits(request_limit=N) |
Manual sliding history window | SlidingWindowConversationManager(window_size=N) |
RunConfig(max_llm_calls=N) |
| Extended thinking | Thinking() capability + anthropic_thinking model settings |
Metrics via run_output.metrics.reasoning_tokens |
Via LiteLLM (provider-dependent) | thoughts_token_count in usage_metadata |
| Command policy | None — model decides | Loop-level CommandPolicy.rewrite() before each MCP call |
BeforeToolCall hook rewrites tool_input in-place |
None — model decides |
| Session abstraction | None | None | None | InMemorySessionService + Runner |
Command policy (
vibepascal_shared.mcp_command_policy) is our own layer — not native to any framework. Agno and Strands use it; Pydantic AI and ADK rely on the model to vary its behavior.
packages/
├── shared/ # vibepascal-shared library
│ ├── pyproject.toml
│ ├── vibepascal_shared/
│ │ ├── guidance_loader.py # load_guidance(), format_guidance_block()
│ │ ├── llm_observability.py # setup_logger(), log_kv(), Timer, …
│ │ └── mcp_command_policy.py
│ ├── client.py # Generic HTTP test client
│ └── mcp_benchmark.py # Multi-client benchmark runner
│
├── pydantic/ # Pydantic AI (2.0.0b3)
│ ├── pyproject.toml
│ └── pydantic_mcp_client.py
│
├── agno/ # Agno 2.6.9+
│ ├── pyproject.toml
│ └── agno_mcp_client.py
│
├── strands/ # Strands Agents SDK + LiteLLM
│ ├── pyproject.toml
│ └── strands_mcp_client.py
│
└── adk/ # Google ADK (MCP)
├── pyproject.toml
└── adk_mcp_client.py
cd packages/<framework>
uv sync [--upgrade] [--prerelease=allow]From the repo root, use the orchestrator scripts:
./pydantic-mcp-game.sh full google:gemini-3.5-flash 1 5
./agno-mcp-game.sh full gpt-4o-mini 1 5
./strands-mcp-game.sh full gemini/gemini-3.5-flash 1 5
./adk-mcp-game.sh full gemini-3.5-flash 1 5Each script automatically invokes uv run --project packages/<fw>, ensuring the right venv is used.
| Dependency | Source | Purpose |
|---|---|---|
python-dotenv>=1.2.1 |
declared | .env loading |
pydantic>=2.0 |
declared (agno, strands, adk) | data validation |
vibepascal-shared |
declared (editable local) | shared logging, guidance, command policy |
httpx>=0.28.1 |
via shared | HTTP client used by shared and most framework SDKs |
asyncio |
stdlib | async runtime — all MCP clients are async def |
logging, os, time |
stdlib | used in every client |
| Framework | Key package | Why isolated |
|---|---|---|
| pydantic | pydantic-ai==2.0.0b3 |
Pre-release; pins pydantic to a pre-release version incompatible with other frameworks. Requires uv sync --prerelease=allow. |
| agno | agno>=2.5.11 |
Bundles its own provider SDKs (anthropic, openai, google-genai, ollama) at potentially different versions than other frameworks need. Also declares mcp>=1.0.0 separately. |
| strands | strands-agents>=1.33.0 + litellm>=1.81.16 |
LiteLLM is a large transitive dependency tree with many provider-specific pins that conflict with direct provider SDKs used by other frameworks. |
| adk | google-adk[extensions,mcp]==2.0.0 |
Pinned version; extensions extra pulls in LiteLLM for non-Gemini model routing. Google-specific and incompatible with agno/strands provider SDK versions. |
agno and strands declare mcp>=1.0.0 explicitly (the reference MCP SDK). pydantic-ai and adk ship their own MCP client implementations — no separate mcp package needed.
- No version conflicts — Each framework can use incompatible versions of shared deps (e.g., pydantic 2.13 vs 2.14a1)
- Smaller venvs — Only install what you need
- Cleaner upgrades — Upgrade one framework without affecting others
- Easier debugging — Dependency issues are scoped to one package
Run all 4 frameworks against a single model:
./play-mcp-game.sh google:gemini-3.5-flash full 1 5This script invokes each *-mcp-game.sh sequentially and compares performance.
- Overview Index: README.md
- Client Implementations: packages/IMPL.md
- Control Flows: packages/FLOW.md
- Observability Setup: packages/shared/OBSERVABILITY.md
- Dependency Rationale: PKG.md — Detailed rationale for system dependencies.