# Split a job across specialist agents with hand-offs

> Split a job across specialist agents with hand-offs: 0 companies fully do it, 19 with a limit, 80 carry the tag. Every listing in Agent frameworks & SDKs with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/frameworks.multi-agent-handoff/
- Markdown: https://www.anchorterminal.com/niches/frameworks.multi-agent-handoff/index.md (~2,100 tokens)
- Slim: https://www.anchorterminal.com/niches/frameworks.multi-agent-handoff/index.min.md (~430 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/frameworks.multi-agent-handoff/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 19 | 80 | 16 | 10 |

## What counts as done

The router sends each test request to the expected specialist, control returns to the caller where the design says it should, each hand-off and the context passed appear as separate steps in the trace, and sub-agent tool calls are visible in the stream. Runs on a second model provider without code changes.

Measured, for our own tests: routing accuracy on a fixed set of 20 requests and the spread of answers across 5 repeats of the same request.

How agents ask for it: route support questions to the right specialist agent; have a manager agent delegate to researchers; build a team of agents for this analysis; hand off to the billing agent and come back

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **AgentOS (Framers Lab, Inc.)**: agency() runs sequential, parallel, debate, review-loop, hierarchical and graph strategies across 13 providers, but the docs list spans for turns and tool results only, not per hand-off, and say nothing of sub-agent tool calls reaching the stream.
- **Agno (Agno Inc.)**: Team modes coordinate, route, broadcast and tasks; members return results to the leader, and tracing records leader coordination and member runs. Member events stream with stream_events. Accounts: tool_call_limit leaves the loop running (agno #8304), and docs were hard to navigate (HN 47561818).
- **BoxLang AI (Ortus Solutions, Corp.)**: Sub-agents are wrapped as delegate_to_<name> tools that return to the parent, across 15+ providers. The docs show no per-hand-off trace span or sub-agent tool calls in the stream, and no OpenTelemetry export.
- **Claude Agent SDK (Anthropic, PBC)**: Subagents run in their own context, in parallel, return a report to the parent and take their own model and maxTurns. Claude models only, so the second-provider test can't be met; per-hand-off trace spans aren't described.
- **CrewAI (CrewAI, Inc.)**: A hierarchical process puts a manager over specialist agents, but accounts report it failing on a local Ollama model where sequential worked (crewAI #1863) and stray backticks behind FastAPI (#2105). Built-in tracing needs a CrewAI AMP login.
- **Cursor Agent SDK (Anysphere, Inc.)**: Subagents declare a description the parent reads to delegate and inherit the model; every model turn runs through Cursor's backend with Cursor models, and the docs read don't show hand-off spans or sub-agent tool calls in the stream.
- **Docker Agent (Docker, Inc.)**: sub_agents delegate through transfer_task and results flow back to the root; handoffs pass the conversation on. --otel spans cover sub-agent hand-offs and provider fallbacks, across many providers. Whether sub-agent tool calls reach the client stream wasn't confirmed.
- **Agent Development Kit (ADK) (Google LLC)**: Sub-agents, workflows and AgentTool, with invoke_agent and execute_tool OpenTelemetry spans. Accounts: control went into a sub-agent and never returned (adk-python #147), a built-in tool failed inside a sub-agent (#53), and some built-in tools exclude all others in one agent.
- **guilds.run (Cigale AI)**: Guilds are standing teams sharing a group chat, and each run keeps the prompt and every tool call. The docs describe chat dispatch between agents rather than a router with a defined return to the caller, and no trace export.
- **LangGraph (LangChain, Inc.)**: Hand-offs are tools returning a Command, with subagent, router and subgraph patterns, on any LangChain model. Tracing is LangSmith, which needs an account, and subgraph hand-offs need careful context handling per the docs.
- **LlamaIndex Framework (LlamaIndex, Inc.)**: AgentWorkflow hands off through can_handoff_to and streams events; the orchestrator pattern exposes sub-agents as tools so control returns. The docs read don't show hand-offs as separate trace spans, and OpenAI is the default LLM unless Settings changes it.
- **Mastra (Kepler Software, Inc.)**: Supervisor agents keep a lead in control and delegate to subagents; hand-off and council patterns are documented, tracing follows model and tool calls in one request, across AI SDK providers. Sub-agent spans per hand-off weren't confirmed; .network() is deprecated.
- **Microsoft Agent Framework (Microsoft Corporation)**: Handoff orchestration passes control between agents with explicit rules (specialists can return to triage) and streams workflow events; OpenTelemetry follows GenAI conventions. An account saw checkpoint resume fail when per-request agent ids changed (agent-framework #4793).
- **OpenAI Agents SDK (OpenAI OpCo, LLC)**: Hand-offs and manager-style Agent.as_tool, whose on_stream callback surfaces nested agent events, and approvals apply across hand-offs. Accounts: hand-offs were one-way (openai-agents-python #847) and routing worsened after a model swap (#1397).
- **Pydantic AI (Pydantic Services Inc.)**: Delegation through tools returns control to the parent, with SubAgents in the Harness, programmatic hand-off and graphs, across 25+ providers, and OpenTelemetry spans via Logfire or any collector. Routing quality is left to the app; no account tested it.
- **smolagents (Hugging Face, Inc.)**: A manager agent calls managed_agents as tools and gets their answers back, on any model through LiteLLM, Transformers or Inference Providers, with OpenTelemetry instrumentation. The secure-execution page says the Docker executor doesn't support multi-agents.
- **Strands Agents (Amazon Web Services, Inc.)**: Agents as tools (control returns to the orchestrator), Swarm hand-offs, Graph and Workflow patterns, with spans per model and tool invocation. Bedrock is the default provider; whether hand-offs get their own span wasn't confirmed.
- **TradingAgents (Tauric Research)**: A fixed LangGraph pipeline of analysts, a bull and bear debate, a trader and risk reviewers, on many providers with a provider per model tier. Routing isn't configurable, and accounts report results changing every run (TradingAgents #178) and suspected training-data leakage in backtests (#33).
- **AI SDK (Vercel Inc.)**: Subagents are invoked through a tool and return a result, optionally streaming progress to the UI, on any AI SDK provider. There is no hand-off primitive; a basic subagent blocks and hides its steps unless you stream it yourself, and spans per hand-off aren't described.

- not this job: NitroStack (Nitrostack Inc.): No agents or hand-offs: a NitroStack project is an MCP server that AI clients connect to.

## What stops agents finishing

- hand-offs that are one-way, so the orchestrator never regains control
- sub-agent run as a tool hides its tool calls from the stream
- built-in tools refused inside sub-agents
- routing quality changes when the default or chosen model changes
- the same input gives different answers each run, so outputs cannot be backtested

## Around it

Before: frameworks.tool-calling-agent. After: agent-observability.debug-failed-run, agent-observability.offline-eval-regression.

## Where this comes from

Weight 16 of 100 in Agent frameworks & SDKs: 10 of 61 accounts have this as their main job (share 16.4, rounded by largest remainder to sum 100); includes 2 TradingAgents accounts on its fixed research pipeline

As data: https://www.anchorterminal.com/api/v1/use-cases/frameworks.multi-agent-handoff.json, or the MCP tool get_use_case with id frameworks.multi-agent-handoff. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
