Capability Negotiation

Overview

Capability negotiation occurs during the handshake. The agent declares which capabilities it wants to use (agent_caps), and the host matches these against its loaded plugins to produce accepted_caps.

Process

1. Agent sends agent_caps list (e.g. ["tools", "memory", "env"])
2. Host looks up each capability in CapabilityRegistry
3. If matching plugins exist → Capability(enabled=True)
4. If no matching plugins → capability is omitted or Capability(enabled=False)
5. Response includes all negotiated results

A2ECapability Enum

All standard capability names:

Enum Value	String	Plugin Type	Description
SKILL	"skill"	SkillPlugin	Sandboxed skill execution
TOOLS	"tools"	ToolPlugin	Primitive tool execution
TOOLKITS	"toolkits"	ToolkitPlugin	Tool bundle management
ENV	"env"	EnvPlugin	RL environment interaction
PROC	"proc"	ProcPlugin	Long-running process management
MEMORY	"memory"	MemoryPlugin	3-tier memory system
LEARNING	"learning"	LearnPlugin	Feedback and adaptation
CHAINS	"chains"	ChainPlugin	DAG pipeline execution
MCP	"mcp"	MCPPlugin	MCP protocol bridge
MULTI_AGENT	"multi_agent"	—	Multi-agent coordination (future)

Negotiation Rules

1. Agent requests, host decides: The host only enables capabilities that have loaded plugins
2. Plugin type mapping: Each plugin declares a type field matching an A2ECapability value
3. Multiple plugins per capability: Multiple plugins can serve the same capability (priority-based dispatch)
4. Exclusive plugins: If a plugin is exclusive=True, it gets sole handling of its message types
5. Priority ordering: Plugins with higher priority are preferred for dispatch ordering

Capability Metadata

Each Capability object in the response carries metadata:

{
  "capability": "tools",
  "enabled": true,
  "metadata": {
    "name": "mytools",
    "type": "tools",
    "priority": 0,
    "exclusive": false
  }
}

After Negotiation

Once capabilities are negotiated:

• The client builds a _capability_map for quick lookup
• The executor's type registry is fully populated
• Only messages for accepted capabilities will be processed
• Messages for rejected capabilities return A2EError with capability_missing

Example

Agent requests:

agent_caps = ["tools", "memory", "env", "chains"]

Host has loaded: ToolPlugin, MemoryPlugin, EnvPlugin (no ChainPlugin)

Result:

accepted_caps = [
    Capability(capability="tools", enabled=True, metadata={...}),
    Capability(capability="memory", enabled=True, metadata={...}),
    Capability(capability="env", enabled=True, metadata={...}),
    Capability(capability="chains", enabled=False, metadata={"reason": "no plugin loaded"})
]