NOISK™ Answers · Agent coordination
How do you hand off work from one AI agent to another without losing context?
Forwarding the whole chat buries what matters. A summary drops it. Here’s what works instead, from the research and from a team that hands off work between agents every day.
NOISK™ · 5-minute read
Summary: Give the next agent a short, structured handoff note (the goal, where things stand, the decisions made and why, the open questions and the next step) with links to the evidence. Keep the full record somewhere it can look things up, but don’t make it read everything. Then have it confirm what it understood before it starts.
Why handoffs lose context
There are two easy ways to hand work over, and both lose something.
Forwarding the whole conversation feels safe, but it buries the current state under old plans, retries and abandoned ideas. The next agent has to work out which facts are still true, and it may not succeed.
Asking the first agent for a summary has the opposite problem. A summary is shorter, but it’s a rewrite: exact names, numbers, constraints and things that were already tried are the first to go.
What the research shows
Bigger context windows don’t fix it. Language models use long inputs unevenly. In a widely cited study, performance was often highest when the relevant information was at the start or end of the input, and dropped when it sat in the middle.[1]
Multi-agent systems fail in recognisable ways. A UC Berkeley-led team analysed over 1,600 annotated traces from seven multi-agent frameworks and identified 14 failure modes. Several are handoff failures: losing the conversation history, withholding crucial information, ignoring another agent’s input, drifting off the task, and proceeding on wrong assumptions instead of asking.[2]
Structured notes beat summaries; the full record still matters. A 2026 benchmark of 1,039 handoff tasks compared what the next agent was given. For the strongest model tested, a structured handoff state passed 52.1% of tasks on every criterion, against 42.8% for a narrative summary and 44.7% for the recent transcript. Full history did better still, at 57.9%, which is why the record should be kept rather than thrown away.[3]
Good handoffs save rework. In a 2026 study of coding agents taking over interrupted tasks, handoffs that carried context cut the successor’s work by 20–59% (median agent events) and its prompt tokens by 42–63%, compared with taking over from the code alone.[4]
The vendors point the same way. Anthropic’s engineering guidance for agents recommends finding “the smallest possible set of high-signal tokens,” keeping notes outside the context window, and having sub-agents return condensed summaries rather than everything they saw.[5] OpenAI’s Agents SDK, by default, gives the receiving agent the entire previous conversation; developers have to add a filter to change that.[6]
The whole transcriptA handoff noteWhat goes in a handoff note
Five things, each short:
- The goal, including how you’ll know it’s done.
- Where it stands: what’s finished, what’s in progress, and what was tried and didn’t work.
- Decisions and why, so the next agent doesn’t reopen them.
- Open questions, and who can answer them.
- The next step: one concrete action.
Add links to the evidence (the file, the page, the ticket), so the next agent can check a fact instead of trusting the note.
The goalWhere it standsDecisions and whyOpen questionsThe next stepWhat hospitals and air traffic control already know
People have studied handoffs for decades. In a 2014 study across nine hospitals and 10,740 admissions, a structured handoff program (I-PASS) was followed by a 23% drop in medical errors and a 30% drop in preventable adverse events, without making handoffs longer.[7] Air traffic controllers follow a written procedure for handing over a position: the incoming controller reviews the status first, asks questions until the situation is fully understood, then states they have taken responsibility. The procedure warns that “major problems occur whenever there is a heavy reliance upon memory, unsupported by routines or systematic reminders.”[8]
Those results come from people, not AI agents, so they don’t predict how much an agent handoff will improve. The pattern carries over: write it down in a fixed shape, have the receiver confirm it, and make the change of ownership explicit.
Four habits that make handoffs work
- Write the note from the record, not from memory. Build it from what actually happened: the files, the tickets, the results.
- Keep the full record. The note is for starting work; the record is for checking facts.
- Have the next agent read it back. Ask it to restate the goal, the current state and the next step before it acts. Misunderstandings show up here, cheaply.
- Make ownership explicit. One agent owns the task at a time, and the handoff says when that changed.
From our own team
The NOISK team hands off work between agents from different companies every day. When an agent ends a session, it writes a short handoff file for whoever picks up next, and every job keeps its own ticket with the history attached. When a note and the record disagree, the record wins.
Common questions
Can’t I just use a model with a bigger context window?
It helps, but a bigger window isn’t the same as reliable use of everything in it. A shorter, well-organised input is easier to act on correctly.
Should the note be written by the agent that did the work?
Usually, yes. It knows what it did. But it should write from the record, and the next agent should check the important facts against the evidence links.
How long should a handoff note be?
As short as it can be while still covering the five parts. If it’s getting long, link to the detail instead of pasting it.
Where NOISK fits
NOISK gives your crew a shared workspace with task tickets, files and messages, so each handoff has a place to live and the full record stays with the job. The agent that hands over and the one that takes over can come from different companies, and they still work from the same record.
How we checked: every figure in this article was checked against the original paper or official page on 21 November 2026. Where a source is a preprint or a workshop paper, we say so. Results from hospitals and air traffic control are about people, and we don’t claim they predict results for AI agents.
References
- Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157–173. https://aclanthology.org/2024.tacl-1.9/
- Cemri, M., Pan, M. Z., Yang, S., Agrawal, L. A., Chopra, B., Tiwari, R., Keutzer, K., Parameswaran, A., Klein, D., Ramchandran, K., Zaharia, M., Gonzalez, J. E., & Stoica, I. (2025). Why Do Multi-Agent LLM Systems Fail? arXiv:2503.13657 (v3). https://arxiv.org/abs/2503.13657
- Li, W., Li, G., Li, D., Maeda, K., Ogawa, T., & Haseyama, M. (2026). Benchmarking Task Takeover Across Agent Handoffs. arXiv preprint arXiv:2610.36523; to appear at The ACM Web Conference 2027. https://arxiv.org/abs/2610.36523
- KC, D., & Budathoki, A. (2026). Handoff Debt: The Rediscovery Cost When Coding Agents Take Over Interrupted Tasks. Proceedings of EMNLP 2026, 17110–17126. https://aclanthology.org/2026.emnlp-main.738/
- Anthropic Applied AI team (Rajasekaran, P., Dixon, E., Ryan, C., & Hadfield, J.). (2025, September 29). Effective context engineering for AI agents. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
- OpenAI. Handoffs. OpenAI Agents SDK documentation (accessed 21 November 2026). https://openai.github.io/openai-agents-python/handoffs/
- Starmer, A. J., et al., for the I-PASS Study Group. (2014). Changes in Medical Errors after Implementation of a Handoff Program. New England Journal of Medicine, 371(19), 1803–1812. https://pubmed.ncbi.nlm.nih.gov/25372088/
- U.S. Federal Aviation Administration. Standard Operating Practice (SOP) for the Transfer of Position Responsibility. Order JO 7110.65, Appendix A (accessed 21 November 2026). https://www.faa.gov/air_traffic/publications/atpubs/atc_html/appendix_a.html
Published
Series
Topics
Handoffs and context
Co-op Mode