BeakrGo to website
Docs/Overview/Agent engine/Meta-tools & handoff

Meta-tools & handoff

Meta-tools let one agent call another. They look like regular tool calls to the LLM, but the system intercepts them and spawns a separate agent that runs to completion before the handoff summary flows back to the parent. The goal is to make multi-agent work structured instead of conversational: outputs are compressed into reusable fields, sources are preserved, and the parent can decide what to do next.

The four patterns cover different coordination shapes. Delegation sends one bounded task to one specialist. Parallel review runs independent specialists at the same time. Pipelines run stages in order when each stage depends on the prior output. Debate asks multiple models or agents to converge on a high-confidence answer.

The four meta-tools

DELEGATEPARALLELPIPELINEDEBATEparentparentparentparentchildone agent,runs to completionABCasyncio.gatherall at oncestage 1stage 2stage 3each stage seesprior output₁₂₃multi-roundconsensus

Same parent, four execution shapes. All four are compiled as child graphs -- the parent sees a single tool result either way.

Signatures

delegate_to_agent
(agent_type, task, audience) -- spawns one child graph, runs to completion, returns the handoff summary as the tool result.
war_room
(agents: [{type, task}, ...]) -- spawns all agents in parallel via asyncio.gather. The user-facing pattern is a parallel specialist review; the internal tool name remains war_room.
agent_pipeline
(stages: [{type, task}, ...]) -- stages run sequentially. Each stage receives the prior stage's output.
debate_ensemble
(question, models) -- multi-round, multi-model debate designed for high-quality consensus.

The handoff protocol

Parent agentChild agent (subgraph)delegate_to_agent(researcher, task)INTERCEPTLoad agent configurationBuild execution graphinject handoff suffix templateSet up communication channelRun child agentLLM LOOP· reasoning · researchemits<handoff>...</handoff><handoff> block parsedPARENT POST-PROCESSstructured fields extractedkey findings -> parent ctxToolResult(handoff_summary)

The parent sees one tool call, one tool result. The child saw an entire ReAct loop. The handoff template is what bridges the two.

The handoff block

When a child agent finishes, it emits a structured <handoff> block. The GraphToolExecutor parses this block and extracts the following fields before returning a ToolResult to the parent:

This avoids agents passing long, messy transcripts back and forth. The parent receives a compact summary, key findings, cited sources, confidence, and gaps, so provenance survives delegation and context stays usable.

SUMMARY
A concise natural-language summary of the child's work -- what was done, what was found, and what conclusions were reached.
KEY_FINDINGS
Bullet-pointed list of the most important facts, data points, or decisions the child produced. These flow into the parent's context window.
SOURCES
References to the tools, documents, or external data the child consulted. Enables provenance tracking through the delegation chain.
CONFIDENCE
The child's self-assessed confidence level in its output -- high, medium, or low -- along with a brief justification.
GAPS
Anything the child could not resolve or areas where additional research would improve the answer. Lets the parent decide whether to delegate again or proceed.