5.c.ii Sub-agent architectures with clean contexts

Hand a context-heavy side task (code search, reading many files, documents or logs) to a sub-agent that starts with a fresh context and only the tools it needs, and have it return a short result or file references so the parent's context stays small. It pays when the parent would otherwise read far more than it can hold: on legal data rooms of up to 80M tokens, delegating bounded reading raised rubric pass rate from 23.3% to 62.4% across seven models. It is not a default win: on bounded repository questions, delegated exploration scored 46.2% against 65.2% for an indexed single agent at 2.3 times the cost per correct answer, with 41.8% of its failures at the hand-off, mostly silent; generic sub-agent roles and a second level of delegation also lost quality.

Tier 4 because it adds sub-agents with their own contexts.

When it applies

In the code: A sub-agent call with a fresh context, a scoped tool list and a summary or file-reference return; or a single agent that reads large search results, file sets, documents or logs into its own context. Warning signs: generic lifecycle sub-agents, sub-agents that inherit the full history or can delegate further, a parent that paraphrases sub-agent output.
In traces: Sub-agent spans: tokens consumed inside vs returned; the parent's peak context; empty returns; facts present in the sub-agent transcript but missing from the parent's answer (hand-off breakdowns, usually silent); share of the corpus read.
In eval results: Single agent vs delegation (and vs an indexed retriever) on search-heavy tasks: success, cost per correct answer.
With a judge: Summary fidelity against the sub-agent's full transcript: facts found but not returned, claims returned but not supported.
With probes: Broad search then synthesis.

How

Start from the single agent: run it on your tasks and log peak context, tokens per task and which tool outputs fill the window; if the material is a corpus you can index, try a retriever before delegating (see 8.a.i), since indexed retrieval beat delegated exploration on bounded repository questions.
Delegate only side tasks whose raw output is large and not needed again (codebase search, reading many files, documents or logs) and whose result fits a short contract; keep work whose steps share a lot of context in the parent.
Define each sub-agent as a task-specific role with an input-output contract: the objective, the inputs it gets, an output schema (findings with file paths, line ranges or document ids, plus what it did not cover) and a length cap; avoid generic lifecycle roles such as requirements, design or code generation.
Start it with a fresh context and a self-contained brief, not the parent's routing messages or whole history, and give it only the tools it needs: read-only search and read tools for exploration, no delegation tool (depth 1), and a turn cap.
Protect the hand-off: have it write long outputs to files and return references (see 6.b.iii), have the parent forward or quote its findings instead of paraphrasing them, and require it to list every policy- or risk-relevant fact it found.
Measure with a paired eval of the single agent against the delegating one on the same tasks and models: success, cost per correct answer, total and parent tokens, the parent's peak context and latency; then label failures where the sub-agent's transcript holds a fact that the parent's answer lacks.

How to measure it

experiment (tier 4): Paired eval on the same tasks and models: the single agent against the same agent with a delegated sub-agent (and against an indexed retriever where the corpus allows): success, cost per correct answer, total tokens, the parent's peak context and latency; code each failure by mechanism, including hand-off breakdowns where the sub-agent found the answer and the parent lost or misreported it.
monitor: Per sub-agent call: tokens consumed inside against tokens returned, empty returns, sub-agent count against task size and the share of the corpus read; on a sample, have a judge check whether the facts in the sub-agent's transcript reach the parent's answer (fact attenuation).

To compare the change against your agent's current version, with the same cases and enough repeats to see past run-to-run noise, use aot-evals.

Expected impact

Quality: mixed, large: Large corpora: +39.1 points rubric pass rate on legal data rooms (Grupen, Sep 2026); folding sub-tasks out of context 0.567 vs 0.540 for a 327K-context agent with the same RL on BrowseComp-Plus, at a 10x smaller active context (Sun, Oct 2025); sub-calls lifted GPT-5 on OOLONG 36.0 to 56.0 but cut Qwen3-Coder on CodeQA 66.0 to 56.0 (Zhang, Dec 2025). Bounded tasks: 46.2% vs 65.2% for an indexed single agent (Oskooei, Aug 2026); task-specific sub-agents +5.91 and +4.46 points, generic ones lower (Hu, Sep 2026); medal rate 45.2% vs 55.7%, CIs overlapping (Brilliantov, Sep 2026); depth 2 cost 19 points (Grupen, Sep 2026)
Cost: mixed, large: More total tokens (Piriyakulkij, Sep 2026): 2.3x cost per correct answer for delegated search (Oskooei, Aug 2026), cost up for six of seven models in a delegating harness while Opus 5 fell from ~$18 to ~$7 per data room (Grupen, Sep 2026); single-agent tokens grew with distractor domains while sub-agent setups stayed flat (Fu-Hinthorn, Jun 2025)
Performance: mixed, small: Each fresh sub-agent must gather its own context; median latency was lower with delegation for two models and higher for two (Oskooei, Aug 2026); RL rollouts slowed from 48 to 74 minutes per step as the trained root sent more sub-agents (Grupen, Sep 2026). Parallel speedups belong to 6.b.iii
Security: mixed, medium (unmeasured): Depends on tool scoping (see 5.c.ii).
Governance: degrades, small (unmeasured): Nested traces.
Brand: none (unmeasured)

Caveats

The large gains come from very large corpora (data rooms, multi-million-token inputs) and from setups that also change other things (a REPL holding the corpus, RL training), so they do not isolate the sub-agent; on bounded tasks with a strong single agent the gain vanished or reversed (repository questions, MLE-bench, generic roles on ProgramBench). The hand-off is the new failure point: it caused 41.8% of delegated-search failures, usually with a confident wrong answer, and a 32B model lost 85% of discovered facts across an orchestrator-subagent boundary, against 3-6% for gpt-4.1-mini. Several studies are small or chart-only (14 data rooms for the depth test, 16-30 facts per cell, SkillsBench accuracies in figures), and Harvey's hold-out sets are unreleased.

Evidence

Evidence comes from: Web research, Document extraction, Retrieval, Code generation, Tool use, Conversation.

Lists this practice in its agent-optimization pattern library (Long-Horizon group)Not measured

Names the practice as a known agent optimization; reports no measured effect for it.

Scale AI, "VeRO: A Harness for Agents to Optimize Agents" (ICML 2026) · Feb 2026 · paper Table 6; pattern library file master-agent-optimization-cookbook.md
Lead agent delegates to subagents with their own context windows on a research task

"a multi-agent system with Claude Opus 4 as the lead agent and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval"

Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, Daniel Ford (Anthropic), "How we built our multi-agent research system" · Jun 2025Older evidence · Benefits of a multi-agent system
Sub-agents work in clean context windows and return a distilled summaryNot measured

"specialized sub-agents can handle focused tasks with clean context windows"; each "might explore extensively, using tens of thousands of tokens or more, but returns only a condensed, distilled summary of its work (often 1,000-2,000 tokens)."

Prithvi Rajasekaran, Ethan Dixon, Carly Ryan, Jeremy Hadfield (Anthropic), "Effective context engineering for AI agents" · Sep 2025Older evidence · Sub-agent architectures
Planner that delegates repository exploration to an isolated sub-agent returning a condensed result, against a single ReAct agent with semantic search over a pre-built indexAgainst

"Semantic search answered 65.2% of questions correctly against 46.2% for deep agentic search"; pass-rate gap 5.6 (Gemini 2.5 Flash), 10.0 (Gemini 2.5 Pro), 10.6 (Qwen3-235B) and 49.6 (Gemini 3 Flash) points, all significant after FDR correction. Pooled cost per correct answer $0.32 vs $0.74, "about 2.3 times more expensive". 68.8% of the delegating agent's tool calls ran inside the sub-agent. "Coordination breakdown accounted for 41.8% of Deep failures", and "in 91% of these cases the agent still returned a fluent, confidently worded answer". Latency "did not order the two paradigms consistently".

Amirkia Rafiei Oskooei, Bora Ilci, Alperen Kayim, Mehmet Egemen Uzun, Berat Can, Kaan Emre Kara, Ozan Orhan, Mehmet S. Aktas, "Deep Agentic Search for Repository-Level Code Question Answering: An Empirical Study" · Aug 2026 · Abstract; §4.1-4.4; §5.1 Table 4; §5.2; §5.3; §6.3-6.4
Single loop vs fixed pipeline vs orchestrator-subagent on policy-governed tasks; count facts a component discovered that never reach the component that must act on themAgainst

Fact attenuation at constraint distance 2, conditional on discovery: Qwen2.5-32B 0/16 (0%) single loop, 9/16 (56%) pipeline, 22/26 (85%) orchestrator-subagent; gpt-4.1-mini 0/27 (0%), 1/30 (3%), 1/18 (6%). The orchestrator-subagent arm found trigger facts in 27 of 100 episodes vs 16 for the single loop, then attenuated 22 of those 27: "The architecture that is better at finding policy-relevant information is worse at carrying it." Dropped facts caused both under- and over-escalation.

Bowen Li, Guojun Wang, "Governance at the Boundary: How Agent Decomposition Degrades Policy Compliance" · Aug 2026 · Abstract; §5.1-5.3, Tables 4-5; §6
Add sub-agents (fresh history, restricted tool whitelist, structured report, nesting disabled) to a minimal coding harness one component at a time: task-specific roles vs generic software-lifecycle rolesMixed

Task-specific sub-agents (behavior analysis, differential testing, submission review) gave "the largest single-component gain of 5.91 percentage points" for Qwen3.7-Max and 4.46 for DeepSeek-V4-Pro. "General subagents also underperform" and "degrade performance for both models". Average prompt tokens: task-specific +7.35% and -9.54%, generic +9.91% and +51.86% (Qwen3.7-Max, DeepSeek-V4-Pro).

Haichuan Hu, Quanjun Zhang, Shengcheng Yu, Zhifei Chen, Tianyu Luo, Chunrong Fang, Zhenyu Chen, Liang Xiao, "Beyond the Model: Demystifying Harness Effects in Software Engineering Agents" · Sep 2026 · §3.1 subagent module; §5.2.1 RQ2.1, Figs 11-12; §5.3.1 Table 7
Enable background sub-agents through the coding agent's native task tool and encourage their use, on top of a single long coding-agent sessionAgainst

With delegation vs base (self-select): percentile 63.62 [59.27, 68.26] vs 66.51 [62.46, 69.97]; medal rate 45.2% [38.1, 54.8] vs 55.7% [48.2, 62.9]. "The base Malena iteration performs no significantly worse than any intervention tested."

Kirill Brilliantov, Alejandro Hernández-Cano, Emmanuel Abbé (EPFL, Apple), "How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering?" · Sep 2026 · §4.4 Multi-agent orchestration, Table 1; Appendix (intervention D)
Run each skill package as a sub-agent in a fresh context instead of loading its instructions into the main agent's contextMixed

With curated packages lacking input-output contracts, "agent skills match or outperform subagents across all models"; with packages that have explicit contracts, "subagents outperform agent skills, with the gain being largest for smaller models". As distracting tools are added, "subagent execution degrades more gracefully". Sub-agents lowered peak context "on over 80% of tasks" for GPT 5.3 Codex and Kimi K2.6, less or reversed for weaker models; "Subagent execution consumes substantially more tokens overall". Routing nodes inline with leaf skills as sub-agents was "consistently best" (Qwen3.5-9B). Accuracies are reported in figures only.

Wasu Top Piriyakulkij, Rachel Lawrence, Alicia Curth, Sushrut Karmalkar, Niranjani Prasad (Cornell, Microsoft Research Cambridge), "Subagents vs Agent Skills: Executing Reusable Knowledge for Long-Horizon Agentic Tasks" · Sep 2026 · §3.1-3.2; §4 Main results, Figs 2-4; Appendix A.1 Fig 5
Agent branches into a sub-trajectory for a sub-task and on return folds it to a short summary, keeping a 32K active context

Untrained, the folding agent beat 32K ReAct and a summary agent (BrowseComp-Plus 0.420 vs 0.286 and 0.386; SWE-Bench Verified 0.492 vs 0.436 and 0.488) but not 327K ReAct (0.478, 0.552). With the same GRPO training: 0.567 vs 0.540 (327K ReAct) and 0.527 (summary) on BrowseComp-Plus; 0.564 vs 0.574 and 0.550 on SWE-Bench Verified. With FoldGRPO: 0.620 and 0.580, "using an active context 10× smaller". Parallel branching (0.6133) performed "similarly to the single-branch version".

Weiwei Sun, Miao Lu, Zhan Ling, Kang Liu, Xuesong Yao, Yiming Yang, Jiecao Chen (ByteDance Seed, CMU, Stanford), "Scaling Long-Horizon LLM Agent via Context-Folding" · Oct 2025Older evidence · Abstract; §4.1 Table 1; §4.5.3
Hold the long input as a variable in a REPL and let the root model call sub-models on slices; compare recursion depth 0 (no sub-calls) with depth 1Mixed

GPT-5 root with GPT-5-mini sub-calls, depth 0 vs depth 1: CodeQA 58.0 vs 62.0, BrowseComp+ (1K) 88.0 vs 91.3, OOLONG 36.0 vs 56.0, OOLONG-Pairs 43.9 vs 58.0. Qwen3-Coder-480B: 66.0 vs 56.0, 46.0 vs 44.7, 43.5 vs 48.0, 17.3 vs 23.1; on CodeQA "the no-sub-calling RLM(depth=0) is able to outperform all sub-calling variants". Sub-calls without offloading the input (CodeAct + sub-calls, GPT-5): OOLONG 40.0, OOLONG-Pairs 28.4. "the median RLM run is cheaper than the median base model run, but more expensive on average due to outlier trajectories".

Alex L. Zhang, Tim Kraska, Omar Khattab, "Recursive Language Models" · Dec 2025 · v3 Table 1; §4 Observations 2 and 4
Root agent with the data room in a REPL delegates bounded reading to sub-agents in their own context windows; only what the root prints enters its context

"Across seven models, the RLM harness raises mean pass rate from 23.3% to 62.4%, a gain of 39.1 percentage points." In the tool loop "no run reads more than 1% of the data room"; with delegation "most read close to all of it". Cost rose for six of seven models; Opus 5 went from ~$18 to ~$7 per data room. The Opus 5 root used 3.8% of input and 1.1% of output tokens. Claude Code and Codex scored 17.9 and 4.6 points below the plain tool loop and "neither spawns sub-agents despite having the ability to". Depth 2 (sub-agents that can delegate) improved 4 of 14 rooms, reduced 10, "lowering mean criteria pass rate by 19 percentage points".

Niko Grupen, Julio Pereyra, Gabe Pereyra, Charlie O'Neill, Mudith Jayasekara, Aaron Ellis-Bloor, Jonathon Liu, Matthew Blau (Harvey), "Post-Training RLM Agents for End-to-End M&A Diligence" · Sep 2026 · Intro; Figs 5-6; Coverage and Cost; Appendix RLM Recursion Depth
Single agent holding every domain's tools and instructions vs supervisor and swarm setups where each sub-agent sees only its own domain, as distractor domains are addedMixed

The single agent "falls off sharply when there are two or more distractor domains" and "uses consistently more tokens" as they grow, while supervisor and swarm "remain flat"; with one distractor domain the single agent did slightly better. Supervisor errors came from "translation" (paraphrasing sub-agent replies); removing hand-off messages from the sub-agent's state, a forward_message tool and tool-name changes gave "a nearly 50% increase in performance". Scores in charts only.

Will Fu-Hinthorn (LangChain), "Benchmarking Multi-Agent Architectures" · Jun 2025Older evidence · Score; Cost (Tokens); Improvements to supervisor

Also known as

Sub-agent spawning with isolated context

In Claude Code

This optimization is a skill in aot-kit, our open-source Claude Code plugin: /aot-kit:isolated-subagent-context. Its library entry holds every source. Have new evidence, or a correction? Suggest a change.

Sub-agent architectures with clean contexts