How henomis/phero works
Phero is a Go framework for building multi-agent AI systems. A typical run starts with an application invoking an agent, which coordinates model calls and tools and can use conversational memory; the application then routes any reported handoffs. The framework also exposes agents remotely through A2A or NATS, and provides a separate RAG path for ingesting and retrieving knowledge. Provider adapters, vector stores, and tracing are supporting layers. Relationships for unsampled capabilities are included only where the README documents them; source-level wiring is cited where excerpts or the source index show it.
Main entry points: agent.go Agent runtime · llm.go LLM interface · processor.go Trace core · vectorstore.go Vector store API
Components 32
Agent orchestration
- Agent runtime
agent/agent.go - Conversation memory
memory/memory.go - Handoff routing
agent/agent.go
LLM services
- LLM interface
llm/llm.go - OpenAI adapter
llm/openai/openai.go - Anthropic adapter
llm/anthropic/anthropic.go - Google adapter
llm/google/convert.go - LLM middleware
llm/middleware/cache.go
Knowledge and memory
- Embeddings
embedding/embedding.go - RAG pipeline
rag/rag.go - Text splitting
textsplitter/textsplitter.go - Vector store API
vectorstore/vectorstore.go - Qdrant store
vectorstore/qdrant/qdrant.go - Postgres vector store
vectorstore/psql/psql.go - Weaviate store
vectorstore/weaviate/weaviate.go - NATS memory
memory/nats/memory.go
Tools and protocols
- A2A server
a2a/server.go - A2A client
a2a/client.go - NATS agent server
nats/server.go - NATS agent client
nats/client.go - MCP adapter
mcp/mcp.go - Skills
skill/skill.go - Skill tool
tool/skill/tool.go - Bash tool
tool/bash/bash.go - Filesystem tools
tool/file/read.go - Web tool
tool/web/web.go - Tool guardrails
tool/permission/permission.go
Tracing
- Trace core
trace/processor.go - OpenTelemetry export
trace/otel/attributes.go - NDJSON export
trace/jsonfile/jsonfile.go - Text trace view
trace/text/text.go
Other components
- Application
Connections 31
- Application → Agent runtime: invokes
- Agent runtime → LLM interface: calls model
- Agent runtime → Conversation memory: uses context
- Agent runtime → Handoff routing: reports handoffs
- Application → Handoff routing: routes next
- LLM interface → OpenAI adapter: implemented by
- LLM interface → Anthropic adapter: implemented by
- LLM interface → Google adapter: implemented by
- LLM interface → LLM middleware: wrapped by
- Agent runtime → Trace core: records spans
- RAG pipeline → Embeddings: embeds text
rag/rag.go - RAG pipeline → Vector store API: queries and writes
rag/rag.go - RAG pipeline → Text splitting: splits documents
- Vector store API → Qdrant store: implemented by
- Vector store API → Postgres vector store: implemented by
- Vector store API → Weaviate store: implemented by
- Agent runtime → NATS memory: uses memory
- A2A server → Agent runtime: exposes
a2a/server.go - A2A client → LLM interface: creates tool
a2a/client.go - NATS agent server → Agent runtime: runs handler
nats/server.go - NATS agent client → LLM interface: creates tool
nats/client.go - MCP adapter → LLM interface: creates tools
mcp/mcp.go - Skill tool → Skills: loads skills
- Agent runtime → Bash tool: runs tool
- Agent runtime → Filesystem tools: runs tool
- Agent runtime → Web tool: runs tool
- Agent runtime → Tool guardrails: guards tools
- Trace core → OpenTelemetry export: exports spans
- Trace core → NDJSON export: writes NDJSON
- Trace core → Text trace view: renders tree
- Trace core → LLM interface: instruments calls