The orchestrator decided. Now what?
In the previous lesson, you learned that a meta-agent — an orchestrator — decides which of your internal agents runs and when. But orchestration only solves the sequencing problem. It determines who goes next. It says nothing about what happens at the boundary between one agent finishing and another one starting.
That boundary is where most of your cognitive value gets destroyed.
You spent forty-five minutes in deep analysis. You identified the structural cause of a recurring problem, weighed three possible interventions, and selected one based on specific evidence. Then your orchestrator handed control to your communication agent — the mode that writes the email, runs the meeting, gives the presentation. And within three minutes of that transition, you could not recall which intervention you selected or why. The analytical agent produced excellent output. The communication agent received a vague impression of "I figured something out." The hand-off failed, and the downstream agent had to either reconstruct the work from scratch or operate on a degraded fragment of what the previous agent knew.
This is not a failure of analysis or communication. It is a failure of context transfer. And it is the single most common way that multi-agent systems — whether human or artificial — lose the value they create.
The cognitive science of context loss
The research on task switching costs quantifies exactly how expensive failed hand-offs are. The American Psychological Association's review of multitasking research found that switching between tasks can cost as much as 40 percent of productive time — not because the tasks themselves are harder, but because the transition between them destroys context. A study from the University of California, Irvine, measured the recovery time: it takes an average of 23 minutes and 15 seconds to fully refocus after an interruption. That 23-minute penalty is not thinking time. It is reconstruction time — the cost of rebuilding the context that was lost in the hand-off.
A 2024 study published in Scientific Reports demonstrated that switch costs scale with the dissimilarity between the tasks being switched. The more different the cognitive demands of the two agents, the more context gets lost in the transition between them. Switching from analytical reasoning to creative writing costs more than switching between two similar analytical tasks. This is because each cognitive mode maintains its own working context — its own set of active representations, priorities, and associations — and these contexts are not automatically portable between modes.
Joshua Rubinstein, David Meyer, and Jeffrey Evans identified two distinct stages of the switch cost in their research on executive control: goal shifting ("I need to do this new thing now") and rule activation ("What are the rules and constraints for this new thing?"). Both stages consume time and cognitive resources, but rule activation is the more expensive — it is where you reload the context that the new agent needs to operate. When the previous agent did not package its output in a way the next agent can use, rule activation takes longer, costs more, and produces more errors.
The average knowledge worker switches tasks more than 300 times per day. Each switch is a hand-off. Each hand-off is a potential context leak. The cumulative cost is not marginal. It is the dominant source of inefficiency in most people's cognitive operations.
Transactive memory: how teams solved this first
The context hand-off problem is not unique to individuals managing internal agents. Teams face the same challenge — and cognitive science has studied their solutions for decades.
Daniel Wegner introduced the concept of transactive memory systems (TMS) in 1985 to describe how groups develop shared mechanisms for encoding, storing, and retrieving information across members. A transactive memory system has three core dimensions: specialization (each member holds distinct knowledge), credibility (members trust each other's expertise), and coordination (members know who knows what and how to access it).
The critical insight for context hand-offs is this: a well-functioning transactive memory system does not require every member to know everything. It requires every member to know what the other members know and how to access that knowledge when they need it. The hand-off protocol is built into the system's structure. When the research specialist finishes their analysis, the team's TMS ensures that the communication specialist knows the analysis exists, knows where to find it, and knows what format it is in.
Research by Lewis (2003) in Organization Science showed that teams with developed transactive memory systems outperform teams without them — not because the individuals are more skilled, but because the transitions between specialized roles lose less context. The TMS functions as a persistent hand-off layer that prevents the team's collective knowledge from degrading every time responsibility shifts from one member to another.
Your internal agents need the same architecture. When your analytical agent finishes, your communication agent should not have to wonder whether analysis was done, what it concluded, or where the output is stored. The hand-off protocol should make all of this explicit and automatic.
SBAR: the hand-off protocol that saves lives
Nowhere are the stakes of context hand-off higher than in medicine. When a nurse finishes a twelve-hour shift and hands a patient's care to the incoming nurse, every piece of context that fails to transfer is a potential medical error. Research published in the British Medical Journal found that 50 percent of hospital staff endorsed the statement that "important patient care information is often lost during shift changes."
The solution that emerged — borrowed from the United States Navy's submarine communication protocols — is SBAR: Situation, Background, Assessment, Recommendation. It is a structured hand-off format that forces the outgoing agent to package context in a way the incoming agent can immediately use.
- Situation: What is happening right now?
- Background: What is the relevant history?
- Assessment: What do I think is going on?
- Recommendation: What should you do next?
When hospitals implemented SBAR and similar structured handover protocols, the results were dramatic. A study at Great Ormond Street Hospital, which adapted Formula 1 pit-stop protocols and aviation handover procedures for surgical-to-ICU transfers, found that technical errors dropped from 5.42 to 3.15 per handover and information omissions dropped from 2.09 to 1.07. A systematic review in 2024 found that the I-PASS handover bundle — another structured protocol — reduced preventable adverse events by 30 percent. The context was the same. The people were the same. The only thing that changed was the structure of the hand-off.
This is the principle: context does not transfer by proximity, by intention, or by shared experience. It transfers by protocol. If you do not have a structured format for packaging what one agent knows into a form the next agent can use, you will lose information at every transition — regardless of how smart, experienced, or well-intentioned the agents are.
The AI parallel: agent hand-offs as engineered context transfer
If context hand-off is a solved problem in medicine and aviation, it is an actively evolving problem in artificial intelligence — and the solutions being developed there illuminate the general principle with unusual clarity.
In multi-agent AI systems, the hand-off problem is literal: when one language model agent finishes its task and passes control to another, the second agent has no memory of what the first one did unless the context is explicitly transferred. Anthropic's engineering team, describing how they built their multi-agent research system, found that "most agent failures are actually orchestration and context-transfer issues." The agents themselves were capable. What failed was the boundary between them.
The OpenAI Agents SDK and LangChain both implement handoffs as first-class primitives — explicit, structured operations that package the outgoing agent's state and transfer it to the incoming agent. In these systems, the hand-off is not a byproduct of the conversation. It is an engineered artifact with a defined schema: the outgoing agent's conclusions, the constraints that must persist, the state variables the incoming agent needs, and the specific action the incoming agent should take first.
Context engineering — treating the information an agent receives as a carefully designed input rather than an accumulated history — has become a core discipline in AI engineering. JetBrains Research published findings in 2025 showing that splitting work among sub-agents, each with their own context window, and then compressing their findings back to a lead agent through structured hand-offs produced far better results than giving a single agent all the context at once. The lesson is counterintuitive: more agents with better hand-offs outperform fewer agents with larger memories. The quality of the transitions matters more than the capacity of any individual agent.
Google's Agent-to-Agent (A2A) protocol and Anthropic's Model Context Protocol (MCP) both standardize the hand-off format between AI agents. These are not suggestions or best practices. They are schemas — structured contracts that specify exactly what information must be packaged, in what format, and with what metadata, every time one agent transfers control to another. The AI engineering community learned through painful experience what hospitals learned a generation earlier: unstructured hand-offs fail. Structured ones succeed. The medium does not matter. The principle is universal.
Your hand-off protocol: building the bridge between agents
The principle is clear. Now make it operational.
Every transition between your internal agents — between analysis and action, between research and writing, between planning and execution, between deep work and communication — needs a hand-off protocol. Not a vague intention to "remember what I was doing." A structured format that the outgoing agent fills in and the incoming agent reads before starting.
Here is a minimal hand-off protocol you can implement today:
State: What did the outgoing agent accomplish? What decisions were made? What conclusions were reached? Write these down in concrete, specific terms — not "I thought about the problem" but "I identified three root causes, ranked them by impact, and selected the first for remediation."
Transfer: What does the incoming agent need to know that it would not otherwise have access to? This includes the reasoning behind decisions, the evidence that was considered and rejected, the constraints that were discovered, and the open questions that remain unresolved.
Action: What should the incoming agent do first? Not "continue the work" but a specific next step: "Draft the stakeholder email starting with the remediation timeline, then address the resource allocation question."
This protocol takes ninety seconds to complete. The context it preserves would otherwise take twenty minutes to reconstruct — if it could be reconstructed at all. Much of what your analytical agent knows is held in working memory associations, emotional valence, and spatial-temporal context that begins decaying the moment attention shifts. The hand-off protocol externalizes this knowledge before it decays.
Write the protocol on paper, in a notes app, or in a dedicated hand-off document. The medium does not matter. What matters is that the context exits the outgoing agent's working memory and enters a stable, accessible format before the transition occurs.
The meta-principle: context is not automatically portable
Beneath the specific techniques — SBAR, State-Transfer-Action, structured AI hand-offs — there is a deeper principle that applies to every multi-agent system, including you.
Context is not automatically portable between agents. It does not transfer by proximity, by intention, or by shared substrate. The fact that your analytical agent and your communication agent share a brain does not mean they share a context. Working memory is modal, attentional, and temporary. What one cognitive mode knows is not available to the next cognitive mode unless it was explicitly externalized.
This is the same principle that makes transactive memory systems work in teams: the system's power comes not from any individual member's knowledge but from the structured protocols that make knowledge transferable between members. It is the same principle that makes SBAR work in hospitals: the patient's condition did not change between shifts, but the incoming nurse's access to relevant information about that condition depends entirely on the quality of the hand-off. And it is the same principle that makes multi-agent AI systems succeed or fail: the agents are only as effective as the transitions between them.
You are a multi-agent system. Your agents are powerful. Your orchestrator, built in the previous lesson, knows which agent should run and when. The question this lesson answers is whether the work one agent does is available to the agent that runs next. Without a hand-off protocol, the answer is: partially, degradedly, and unreliably. With one, the answer is: completely, structurally, and by design.
From hand-off to dependency map
You now have the ability to orchestrate your agents (L-0508) and transfer context between them (this lesson). But which agents need context from which other agents? Which hand-offs are critical and which are optional? Where are the transitions that, if they fail, cascade into system-wide breakdowns?
These are dependency questions, and they require a map. In the next lesson (L-0510), you will learn to draw the dependency graph between your agents — making visible not just the hand-offs that exist but the ones that are missing. The hand-off protocol gives you the mechanism. The dependency map shows you where to deploy it.
Sources:
- Rubinstein, J. S., Meyer, D. E., & Evans, J. E. (2001). "Executive Control of Cognitive Processes in Task Switching." Journal of Experimental Psychology: Human Perception and Performance, 27(4), 763-797.
- Mark, G., Gudith, D., & Klocke, U. (2008). "The Cost of Interrupted Work: More Speed and Stress." Proceedings of CHI 2008, ACM.
- Wegner, D. M. (1985). "Transactive Memory: A Contemporary Analysis of the Group Mind." In B. Mullen & G. R. Goethals (Eds.), Theories of Group Behavior. Springer.
- Lewis, K. (2003). "Measuring Transactive Memory Systems in the Field." Journal of Applied Psychology, 88(4), 587-604.
- Catchpole, K. R., et al. (2007). "Patient Handover from Surgery to Intensive Care: Using Formula 1 Pit-Stop and Aviation Models to Improve Safety and Quality." BMJ Quality & Safety, 16(1), 44-48.
- Starmer, A. J., et al. (2014). "Changes in Medical Errors after Implementation of a Handoff Program." New England Journal of Medicine, 371(19), 1803-1812.
- Anthropic. (2025). "How We Built Our Multi-Agent Research System." Anthropic Engineering Blog.