Position Effects and Structured Subagent Outputs

1.The lost-in-the-middle effect

Context rot describes recall dropping as total length grows. The lost-in-the-middle effect adds that the drop is uneven across positions. Anthropic Academy compares it to the human serial position effect: items at the start of a list benefit from primacy and items at the end benefit from recency. The Academy page reports that language models show the same pattern. In a 2023 Stanford study, accuracy was highest when the key fact sat at the very beginning or very end of a long context, and fell by more than 30% when the fact was in the middle. The page attributes this to how attention works: transformer attention gives more weight to the edges of the context window.

This matters most when an agent aggregates results. A coordinator that pastes eight subagent reports one after another has put reports three to six in the weakest positions. A finding that appears only in report five is the one most likely to be left out of the final synthesis, even when everything fits easily in the window.

Sources1

2.Laying out aggregated inputs to work with position

You can't get rid of position effects, but you can choose what goes in the strong positions. Start the aggregated input with a short summary of the key findings, so the conclusions most likely to be needed sit where attention is strongest. Then put the detailed results under explicit section headers, one per source or subtask, so each finding is clearly separated rather than lost in continuous text.

Anthropic's prompting guidance supports this layout for inputs over about 20k tokens. Put the long material near the top, above the query and instructions. Wrap each document in its own tags with a source subtag and any other metadata. For long-document tasks, ask Claude to quote the relevant passages before answering, which helps it focus on the parts that matter. Note that this guidance puts the query at the end, after the documents, so a findings summary at the top and the question at the bottom fit together.

<documents> <document index="1"> <source>annual_report_2023.pdf</source> <document_content> {{ANNUAL_REPORT}} </document_content> </document>

Sources2

3.Subagent outputs must carry their metadata

A coordinator can only synthesize accurately if it can tell where each claim came from and how old it is. When a subagent returns a paragraph of conclusions with no dates or sources, the coordinator has nothing to go on when two subagents disagree. It can't tell which price is current or which source is more authoritative. The fix belongs in the delegation instructions. Anthropic's multi-agent research write-up says every subagent needs an objective, an output format, guidance on tools and sources, and clear task boundaries. The output format should require metadata alongside each finding: the date of the information, the source location, and the method used to get it.

This is the same principle as the source tags in the multi-document prompt structure, applied to agent-to-agent handoffs. The research system goes further at the end: a dedicated citation agent matches each claim in the report to a specific location in the sources. That step only works if the source information made it through every handoff before it.

Sources32

4.Return structured data, not reasoning chains

The other half of the subagent contract is size. A subagent may use tens of thousands of tokens exploring a problem, but in Anthropic's description it returns only a condensed summary, often 1,000 to 2,000 tokens. That keeps the detailed search work inside the subagent and the coordinator's context free for synthesis. If a downstream agent has a limited budget, change the upstream agents to return key facts and citations as structured data instead of their full reasoning and verbose content. The narrative of how a subagent got there costs the coordinator tokens and adds nothing to the synthesis. (The exam guide also lists relevance scores as an output field. None of the sources here discuss them specifically.)

For large artifacts, the multi-agent research system goes a step further. Subagents save their output to external storage and pass lightweight references back to the coordinator. This avoids the token cost of copying large outputs through conversation history and reduces information loss across stages. In all of these cases the goal is the same: make each token that reaches the downstream agent carry information it needs.

Sources435

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.If every subagent report fits in the context window, the model will weigh all of them equally.Why is that wrong?

   Fitting in the window doesn't mean getting equal attention. Content in the middle of a long input is the most likely to be missed.

   Covered in The lost-in-the-middle effect

2. 2.Subagents should return only clean conclusions. Dates and sources are clutter the coordinator doesn't need.Why is that wrong?

   Without source and date metadata the coordinator can't settle conflicting findings, and claims can't be attributed to their sources.

   Covered in Subagent outputs must carry their metadata

Practise it for real

Turn on tool result clearing for a tool-heavy agent while protecting the tool that holds case facts

1. 1.Send a request with context_management={"edits": [{"type": "clear_tool_uses_20250919"}]} and the context-management-2025-06-27 beta header

   Why: With only the type set, clearing starts at the default trigger of 100,000 input tokens and keeps the last 3 tool uses

   You should see: A normal response. Clearing only starts once the prompt passes the trigger

2. 2.Add trigger, keep and exclude_tools to the edit, listing your case-facts or memory tool in exclude_tools

   Why: exclude_tools ensures the results your facts depend on are never cleared

   You should see: In a long run, old results from other tools are replaced with placeholder text, and the excluded tool's results stay in full

3. 3.Compare the usage field on responses before and after clearing starts

   Why: Every response reports how many tokens the request used

   You should see: Input token counts drop once clearing starts, while your local history stays complete