Lost in Delegation — Delegation vs direct tool use — What does it mean?
The news. On November 12, 2026, researchers at Ruhr University Bochum posted Lost in Delegation (arXiv 2611.16526). They built four fact-lookup datasets (book release years, ages of Nobel laureates, GDP growth rates, product launch years), each split into facts from before and after the models' knowledge cutoffs. Four models (Claude Sonnet 4.5, GPT-5, gpt-oss-120b, Qwen3-30B-A3B-Thinking-2507) ran in four setups: no tool, a single agent with web search, forced delegation and optional delegation. Orchestrator and executor are always the same model, and both delegation setups use the same search tool as the single agent. Read the paper →
Picture an editor who needs five release years. One way is to walk down to the archive and look up each title: the desk fills with papers, but every dead end happens in front of the editor. The other way is to phone a researcher for each title and write down what comes back. The desk stays clean, and that is the whole case for the orchestrator-workers pattern in search: each lookup runs in its own short context, which the curriculum calls Subagents as Context Isolation.
On these tasks, the phone line came with an accuracy cost: optional delegation was worse than the single agent in 22 of 32 combinations, and forced delegation in 21. Nothing else changed between the setups. The model, the search tool and the questions were the same; only the question of who holds the tool was different.
The first loss is in the note. When the editor dictates a request, it often adds what it thinks it knows, such as the author, or whether the title is a novel or a film. For a book the model already knows, that helps the search. For a book published after its knowledge cutoff, the guesses go wrong: under delegation, no author name GPT-5 added to a post-cutoff book query was correct, and only 4–6% of the medium labels it added (novel, film, song) were. Asked for the 2026 novel The Night We Met, GPT-5 sent the executor to look up the release year of the Lord Huron song with that title. The executor correctly answered 2015, which is a right answer to the wrong question. A single agent can put the same wrong guess into its own search, but it still reads the results, so it has a chance to notice the mismatch.
The second loss is on the way back. Only the executor's final answer crosses the boundary, not the searches it ran, the pages it opened, or the locked doors it hit. Locked doors are common: in the single-agent setup, 13% of page visits returned a 403 error, and under delegation the rate was 17% (forced) and 20% (optional). When the error stays in the agent's own context, the agent can try another source. When it happens on the far side of the boundary, the orchestrator often never learns about it. In one GDP trace, the executor could not open a World Bank page because the site blocks automated clients, did not report it, and the orchestrator kept sending requests that hit the same block. This is the Handoffs problem in its plainest form: a handoff carries only what the sender writes down.
The third loss is trust. An answer that arrives without its evidence is easy to doubt. In one trace, the executor correctly reported that A Gathering of Shadows came out in 2016, and the orchestrator rejected it in favour of its own wrong memory of 2022. Other models doubt by asking again: GPT-5 re-delegated up to 54% of tasks on pre-cutoff GDP, and re-delegation stayed common even when the executor's answer was already right. Every repeat is one more call, which is the Coordination Costs side of a team design.
Where the accuracy goes (averaged over the four models and four datasets, the paper's Table 1). Hold the model, the search tool and the question format fixed, and change only the cutoff split. The single agent goes from 0.77 to 0.75, a 2-point drop, because it searches for what it does not know. Forced delegation goes from 0.77 to 0.66 and optional delegation from 0.74 to 0.64, drops of 11 and 10 points. So before the cutoff, forced delegation ties the single agent; after the cutoff the gap is 9 percentage points (0.75 vs 0.66), which by our arithmetic is about 9 more wrong answers per 100 items. The worst single case is GPT-5 on post-cutoff GDP: 0.89 with direct search, 0.35 under forced delegation and 0.43 under optional delegation. The locked pages show the same pattern: on items where a page visit hit a 403, the single agent still answered about 73% correctly, forced delegation 54% and optional delegation 52%. By our arithmetic, that is about 19 more wrong answers per 100 such items once the error sits behind the boundary.
| Setup (average of 4 models × 4 datasets) | Pre-cutoff accuracy | Post-cutoff accuracy | Change | Source |
|---|---|---|---|---|
| No tools (closed book) | 0.53 | 0.01 | −98% | Table 1 |
| Single agent (direct web search) | 0.77 | 0.75 | −3% | Table 1 |
| Optional delegation | 0.74 | 0.64 | −14% | Table 1 |
| Forced delegation | 0.77 | 0.66 | −13% | Table 1 |
Can a better prompt repair the line? The authors tried two fixes. Telling agents not to repeat equivalent searches barely changed how similar the follow-up searches were. Telling executors to report inaccessible pages raised GPT-5's GDP accuracy under both delegation modes but left it well below the single agent, and it lowered gpt-oss-120b's accuracy. The structural fix, sending the executor's full trace back, would restore the evidence but also undo the context saving that was the reason to delegate. So the question in When NOT to Use an Agent has a sibling: add a subagent hop only when the subtask gains more from a clean context than it loses at the handoff. Two limits apply. The tasks are small sets of independent facts, and orchestrator and executor are always the same model; the paper does not test mixed-model teams or subtasks that depend on each other.
Goes deeper in: Agent Engineering → Agent Teams → Handoffs
Related explainers
- Cohere coding-agent teams — Planning-only delegation — the opposite result: a team that delegates one stage, not every stage, and pays off.
- MSR delegation study — Cascading fidelity loss over 20 iterations — what happens when a handoff is repeated many times.
- BLADE — Handoff auditing — how to find which handoff introduced an error.
- Deep Agents — Forked vs isolated subagent context — starting a subagent from a copy of the supervisor's context instead of an empty one.