AgentPool Orchestration Hub
Configure heterogeneous AI executors in YAML and expose them through common agent and server protocols.
Per-dimension scores and reasoning
Evidence: The project provides an ACP server with tool confirmations (README mentions 'bidirectional communication with tool confirmations'), but no detailed permission model or least privilege principle. Many dependencies, but no security audit or vulnerability scanning evidence. External effects include starting servers, executing commands, but no explicit restrictions. Rollback mechanism not mentioned. Source attribution: author info clear, but publisher unverified. Deductions: lack of specific security mechanism descriptions, opaque permission control, missing rollback.
Evidence: README and pyproject.toml descriptions are consistent, version number clear, but no test results or CI status provided. Many dependencies, some self-maintained, but no availability guarantees. Failure messages not documented. Deductions: dependency availability unverified, failure handling documentation insufficient.
Evidence: Targeted at developers, provides multiple usage scenarios (CLI, servers, programmatic), supports multiple agent types and protocols. Capability boundaries defined via YAML config, but not explicitly limited. Triggers mentioned but details insufficient. Environment fit: supports multiple platforms (Linux, macOS, Windows), but requires Python >=3.13. Deductions: trigger precision and boundary descriptions insufficient.
Evidence: README well-structured, provides quick start, config examples, feature list. Install instructions simple (uv tool install). Naming stable (agentpool). Examples abundant, but no FAQ. Known limitations not explicitly listed. License MIT, version clear, but no changelog. Maintenance responsibility: author clear, but no maintenance policy. Deductions: missing FAQ, known limitations, and changelog.
Evidence: Output usability high, provides multiple output formats (structured output, streaming). Marginal value: solves multi-agent coordination problem, provides unified interface. Cost-benefit: many dependencies, complex installation, but powerful. Deductions: cost-benefit not quantified, many dependencies may increase maintenance cost.
Evidence: README claims (e.g., support for multiple protocols) partially correspond to pyproject.toml dependencies and entry points, but no independent verification. Cross-source corroboration: documentation links exist, but no test results or third-party reviews. Fact-inference separation: README distinguishes feature descriptions and examples, but does not explicitly mark inferences. Deductions: lack of independent verification and test evidence.
- Publisher identity unverified; assess risks before use.
- Many dependencies, some self-maintained, potential supply chain risks.
- No security audit or vulnerability scanning evidence; permission control details opaque.
- No rollback mechanism; configuration changes may be hard to revert.
- Requires Python >=3.13, may limit deployment environments.
What does this agent do, and when should you use it?
AgentPool is a YAML-centered orchestration project for native PydanticAI agents, Claude Code, Codex, ACP agents, and AG-UI agents. Its unified interface supports delegation, message routing, shared context, parallel teams, and sequential chains. It runs as a Python library and provides CLI entry points plus ACP, OpenCode, MCP, AG-UI, and OpenAI API servers. Native agents can run prompts, stream events, and accept multimodal input, while configuration can define models, tools, MCP servers, knowledge paths, connections, triggers, and storage. It fits engineering teams that need one control surface for several existing agent systems and client protocols.
You declare agents in agents.yml with a type, model, tools, and connections, using types such as native, claude_code, codex, acp, and agui. AgentPool then runs a named entry with agentpool run assistant "Hello!", or exposes configured agents with agentpool serve-acp, serve-opencode, serve-mcp, serve-agui, or serve-api. In Python, async with AgentPool("agents.yml") as pool loads the configuration; pool.get_agent("assistant") returns an agent that can be called with await agent.run(...) or agent.run_stream(...). Teams can run in parallel with &, pipelines can run sequentially with |, and a native coordinator using the subagent tool can delegate to other agents in the pool.
- An IDE-focused developer using Zed, Toad, or another ACP client who wants configured agents behind an ACP server with tool confirmations and session management.
- An engineering team using Claude Code, Codex, and Goose that wants a native coordinator to delegate work across those executors.
- An OpenCode TUI/Desktop user who needs serve-opencode to control configured agents, including agents operating with local or fsspec remote filesystems.
- A team building an AG-UI frontend that needs to place a URL-backed AG-UI agent alongside other configured executors.
- A team with callers expecting an OpenAI-compatible endpoint that wants to expose the pool through serve-api.
What are this agent's strengths and limitations?
- A single YAML configuration can describe native PydanticAI, Claude Code, Codex, ACP, and AG-UI types.
- It offers ACP, OpenCode, MCP, AG-UI, and OpenAI API server boundaries for IDE, TUI, frontend, and API use cases.
- Parallel teams, sequential chains, and subagent-based delegation have explicit configuration or API forms.
- Documented capabilities include streaming, multimodal input, structured output, triggers, and local/remote file abstraction.
- The examples do not specify a Python version, provider credential setup, authentication flow, or server port configuration.
- External agents and models depend on their respective runtimes, availability, and protocol compatibility; failure behavior is not described.
- The broad YAML surface spans models, tools, connections, knowledge, and storage, creating configuration design and maintenance work.
- OpenCode remote filesystem support relies on fsspec, but supported backends, permission handling, and isolation are not documented.
How do you install or deploy this agent?
The documented installation is:
uv tool install agentpoolCreate agents.yml:
agents:
assistant:
type: native
model: openai:gpt-4o
system_prompt: "You are a helpful assistant."The supplied material does not document a Python version, provider credential setup, or environment-variable names. This example uses an OpenAI model, so access credentials required by the chosen model provider must be available before running it.
How do you use this agent?
After creating agents.yml, run:
agentpool run assistant "Hello!"To start an ACP server:
agentpool serve-acp agents.ymlPython usage:
from agentpool import AgentPoolasync with AgentPool("agents.yml") as pool:
agent = pool.get_agent("assistant")
result = await agent.run("Hello")Other documented commands include agentpool serve-opencode config.yml, agentpool serve-mcp config.yml, agentpool watch --config agents.yml, and agentpool history stats --group-by model.
How does this agent compare with similar options?
AgentPool is positioned as a protocol bridge rather than a replacement for Claude Code, Codex, Goose, or a standalone AG-UI agent: those can be configured as different pool types and reached through common ACP, OpenCode, or AG-UI interfaces.