copass-anthropic-agents 0.2.0

Copass agent primitives for the Anthropic Managed Agents API

copass-anthropic-agents

Anthropic Managed Agents backend for Copass. Depends on copass-core-agents, which owns the vendor-neutral primitives (BaseAgent, AgentTool, AgentBackend ABC, events, AgentScope, registries, etc.) shared across every provider-specific Copass SDK.

This package owns the Anthropic-specific piece — ManagedAgentBackend and the CopassManagedAgent convenience subclass — and re-exports the core primitives for one-line dev imports. Either of these works for consumers:

from copass_anthropic_agents import BaseAgent, CopassManagedAgent  # convenient
# or
from copass_core_agents import BaseAgent                           # explicit boundary
from copass_anthropic_agents import CopassManagedAgent

Pairs with Copass's server-side router — both sides import the same code so the agent runtime is a single codepath, whether the agent runs in the dev's process (deep-integration) or server-side via Copass's Router endpoint (thin client).

Install

pip install copass-anthropic-agents

Requires anthropic>=0.93.0 (first version shipping the beta.agents / beta.environments / beta.sessions surface).

Quickstart

import os
from copass_anthropic_agents import (
    AgentInvocationContext,
    AgentScope,
    CopassManagedAgent,
)

agent = CopassManagedAgent(
    identity="support",
    system_prompt="You are a support agent.",
    anthropic_api_key=os.environ["ANTHROPIC_API_KEY"],
)

ctx = AgentInvocationContext(scope=AgentScope(user_id="u-123"))
result = await agent.run(
    messages=[{"role": "user", "content": "Hello"}],
    context=ctx,
)
print(result.final_text)

Streaming

from copass_anthropic_agents import (
    AgentFinish,
    AgentTextDelta,
    AgentToolCall,
    AgentToolResult,
)

async for evt in agent.stream(messages, context=ctx):
    match evt:
        case AgentTextDelta(text=chunk):
            print(chunk, end="", flush=True)
        case AgentToolCall(name=name, arguments=args):
            print(f"\n[tool-call] {name}")
        case AgentToolResult(name=name, result=result):
            print(f"\n[tool-result] {name}")
        case AgentFinish(stop_reason=reason, session_id=sid):
            print(f"\n[finish] {reason} session={sid}")

Multi-turn continuation

The Anthropic managed-agent session holds conversation state. Capture AgentFinish.session_id and thread it back on the next turn:

from copass_anthropic_agents import SESSION_ID_HANDLE

ctx2 = AgentInvocationContext(
    scope=AgentScope(user_id="u-123"),
    handles={SESSION_ID_HANDLE: prior_session_id},
)

Custom tools

See AgentTool + AgentToolRegistry + AgentToolResolver. Tools are dispatched locally on tool-use events; provider handles the rest.

from copass_anthropic_agents import AgentTool, AgentToolRegistry, ToolSpec

class EchoTool(AgentTool):
    @property
    def spec(self):
        return ToolSpec(
            name="echo",
            description="Echo a message back.",
            input_schema={
                "type": "object",
                "properties": {"text": {"type": "string"}},
                "required": ["text"],
            },
        )

    async def invoke(self, arguments, *, context=None):
        return {"echoed": arguments["text"]}


tools = AgentToolRegistry()
tools.add(EchoTool())
agent = CopassManagedAgent(
    identity="support",
    system_prompt="You are helpful.",
    anthropic_api_key=os.environ["ANTHROPIC_API_KEY"],
    tools=tools,
)

Rate-limit note

ManagedAgentBackend caches Anthropic-side agent + environment resources per-instance by a fingerprint of (model, system_prompt, tool_schemas). Reusing the backend across requests shares the cache; constructing a fresh backend per request re-creates resources and hits Anthropic's 60-creates/min limit fast. Use the backend as a process-wide singleton in server contexts.

Context injection (deferred)

A future release will pull Copass's Context Window into system_prompt via the reserved {{copass_context}} placeholder (see spec/context-placeholders.md in this monorepo). Not wired in this release — injection would change system_prompt per invocation, invalidating the backend's fingerprint cache.

License

MIT.