ContextRender: From Execution Dependencies to Agent Context

Savini Kashmira; Jayanaka L. Dantanarayana; Lingjia Tang; Jason Mars

cs.AISep 29, 2026

ContextRender: From Execution Dependencies to Agent Context

Authors: Savini Kashmira, Jayanaka L. Dantanarayana, Lingjia Tang, Jason Mars

Organizations: University of Michigan, Ann Arbor, USA

Abstract

LLM agents performing long-horizon tasks accumulate tool results that later steps may need. Passing the full history to every invocation is costly even when it fits within the context window, while reducing it risks omitting needed information. Existing context management methods can overlook how earlier tool results are used in subsequent execution, leaving needed information out of context. We introduce ContextRender, which manages context through a persistent graph of execution dependencies. We develop Tool-Flow Analysis to track how later operations reuse information from earlier tool results, providing a signal called observed reuse. A renderer combines this signal with recency and semantic relevance to select results within a fixed history budget, retaining omitted results for later use. Across AppWorld and 8-objective QA with three execution models, ContextRender outperforms the evaluated context management baselines using a 6K history budget, well below the models' maximum context windows. Within this budget, it achieves task performance close to or above that of passing the full history while reducing mean inference cost by 10.2%-32.2% relative to Full history. Ablations show that observed reuse improves task performance and retention of results reused later.

Figures & tables

Figure 1: Comparison of context management approaches. ContextRender selects history from a persistent graph using recency, semantic relevance, and observed reuse, retaining unselected results for later use.

Figure 2: Execution trace from an AppWorld task in which the agent settles a shared dinner bill among roommates. Information from the contact result produced at Turn 8 is used by later operations at Turns 9, 15, and 16. At Turn 18, the agent needs Nancy’s email address, which has not been copied forward and remains available only from the Turn 8 result.

Figure 3: ContextRender architecture. The runtime adapter records messages, tool calls, and complete tool results in a persistent context graph. Tool-Flow Analysis records observed reuse as execution dependencies. Before each model invocation, the renderer includes the accumulated execution history if it fits within the history budget $B$ ; otherwise, it selects tool results using observed reuse, recency, and semantic relevance. Omitted results remain stored for later invocations.

Figure 4: Overall performance (AppWorld TGC; QA F1) across all tasks versus mean standardized inference cost on the same tasks that trigger context management at

B=6

K. Costs are normalized to the mean cost of Full history (

=1

). “Recency” denotes Recency + Summary.

Model	Benchmark	Full history		Prune		Compact		Recency + Summary		Semantic Retrieval		ACON
		Perf.	Cost	Perf.	Cost	Perf.	Cost	Perf.	Cost	Perf.	Cost	Perf.	Cost
GPT-4.1	AppWorld Normal	$\downarrow$ 2.3%	$\downarrow$ 32.2%	$\uparrow$ 14.4%	$\downarrow$ 3.2%	$\uparrow$ 12.4%	$\downarrow$ 0.8%	$\uparrow$ 7.6%	$\downarrow$ 1.6%	$\uparrow$ 7.6%	$\downarrow$ 18.4%	$\uparrow$ 3.3%	$\downarrow$ 14.9%
	AppWorld Challenge	$\uparrow$ 8.5%	$\downarrow$ 28.8%	$\uparrow$ 19.7%	$\downarrow$ 2.7%	$\uparrow$ 12.1%	$\downarrow$ 1.8%	$\uparrow$ 11.6%	$\downarrow$ 16.7%	$\uparrow$ 14.4%	$\downarrow$ 9.7%	$\uparrow$ 11.1%	$\downarrow$ 8.1%
	8-objective QA	$\uparrow$ 4.5%	$\downarrow$ 10.2%	$\uparrow$ 3.7%	$\downarrow$ 2.0%	$\uparrow$ 2.1%	$\downarrow$ 4.5%	$\uparrow$ 9.5%	$\rightarrow$ 0.0%	$\uparrow$ 14.4%	$\rightarrow$ 0.0%	$\uparrow$ 10.4%	$\downarrow$ 5.7%
Muse Glimmer 30B	AppWorld Normal	$\uparrow$ 5.8%	$\downarrow$ 12.4%	$\uparrow$ 14.2%	$\uparrow$ 1.7%	$\uparrow$ 7.4%	$\uparrow$ 0.8%	$\uparrow$ 8.2%	$\downarrow$ 3.2%	$\uparrow$ 10.7%	$\downarrow$ 11.1%	$\uparrow$ 9.8%	$\downarrow$ 2.4%
	AppWorld Challenge	$\uparrow$ 2.9%	$\downarrow$ 19.9%	$\uparrow$ 16.8%	$\downarrow$ 7.2%	$\uparrow$ 9.2%	$\downarrow$ 16.8%	$\uparrow$ 10.1%	$\downarrow$ 2.3%	$\uparrow$ 13.6%	$\downarrow$ 8.5%	$\uparrow$ 8.2%	$\downarrow$ 12.2%

Table 1: Relative changes in task performance and mean inference cost of ContextRender versus each baseline and Full history . Green, red, and gray arrows indicate improvements, regressions, and no change, respectively.

Variant	Performance		Later-Used Result Recall		AppWorld Later-Used Result Recall by Result Age
Variant	AppWorld (TGC)	QA (F1)	AppWorld	QA	$\mathbf{\leq 3}$	4–10	11–25	$\mathbf{>25}$
CR–Recency	71.4	48.5	63.8	86.0	95.7	78.7	37.7	18.5
CR–Relevance	73.2	47.5	71.4	64.2	89.5	81.9	54.1	47.1
CR–Recency + Relevance	72.6	50.8	61.6	76.1	84.4	72.5	43.9	22.7
CR–Full (+Reuse)	75.6	53.1	87.2	91.0	99.3	91.5	79.0	68.5

Table 2: Ablation of ContextRender with GPT-4.1 on AppWorld Test-Normal and 8-objective QA at

B=6

K. Recall denotes Later-Used Result Recall. Result age is the number of turns between a tool result’s creation and its subsequent reuse. All values are percentages.

Figure 5: Budget sensitivity with GPT-4.1.

B

denotes the history budget. “Recency” denotes Recency + Summary.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Selector	$W_{\mathrm{rec}}$	$W_{\mathrm{rel}}$	$W_{\mathrm{reuse}}$
CR–Recency	1.0	0.0	0.0
CR–Relevance	0.0	1.0	0.0
CR–Recency + Relevance	1.0	1.0	0.0
CR–Full (+Reuse)	1.0	1.0	1.0

Appendix

Table 3: Scoring weights used in the selector ablations. Zero disables the corresponding term; retained weights remain unchanged.

Result age at reuse	Events
$\leq 3$ turns	713
4–10 turns	1,124
11–25 turns	1,099
$>25$ turns	238
Total	3,174

Appendix

Table 4: AppWorld reuse events by result age.

Result age at reuse	Events
$\leq 3$ turns	224
4–10 turns	70
11–25 turns	41
$>25$ turns	0
Total	335

Appendix

Table 5: 8-objective QA reuse events by result age.

Method	AppWorld Test-Normal			AppWorld Test-Challenge			8-objective QA
Method	TGC (%)	Cost	Norm.	TGC (%)	Cost	Norm.	F1 (%)	Cost	Norm.
Full history	77.4	0.177	1.000	51.1	0.302	1.000	50.8	0.167	1.000
Prune	66.1	0.124	0.701	46.3	0.221	0.732	51.2	0.153	0.916
Compact (LLM)	67.3	0.121	0.684	49.4	0.219	0.725	52.0	0.157	0.940
Recency + Summary	70.2	0.122	0.689	49.6	0.258	0.854	48.5	0.150	0.898
Semantic Retrieval	70.2	0.147	0.831	48.4	0.238	0.788	46.4	0.150	0.898

Appendix

Table 6: Overall task performance and mean inference cost for GPT-4.1. Performance is measured across all tasks; cost is measured on the same tasks that trigger context management when accumulated execution history exceeds

B=6

K. Cost is in reference USD; normalized cost uses uncapped Full history (

=1

) on the same tasks in each benchmark.

Method	AppWorld Test-Normal			AppWorld Test-Challenge			8-objective QA
Method	TGC (%)	Cost	Norm.	TGC (%)	Cost	Norm.	F1 (%)	Cost	Norm.
Full history	81.5	0.137	1.000	58.3	0.161	1.000	50.4	0.269	1.000
Prune	75.6	0.118	0.861	51.3	0.139	0.863	49.3	0.243	0.903
Compact (LLM)	80.4	0.119	0.869	54.9	0.155	0.963	50.4	0.259	0.963
Recency + Summary	79.8	0.124	0.905	54.4	0.132	0.820	48.2	0.245	0.911
Semantic Retrieval	78.0	0.135	0.985	52.8	0.141	0.876	51.1	0.235	0.874

Appendix

Table 7: Overall task performance and mean inference cost for Muse Glimmer 30B. Performance is measured across all tasks; cost is measured on the same tasks that trigger context management when accumulated execution history exceeds

B=6

K. Cost is in reference USD; normalized cost uses uncapped Full history (

=1

) on the same tasks in each benchmark.

Method	AppWorld Test-Normal			AppWorld Test-Challenge			8-objective QA
Method	TGC (%)	Cost	Norm.	TGC (%)	Cost	Norm.	F1 (%)	Cost	Norm.
Full history	82.7	0.1425	1.000	77.0	0.236	1.000	56.5	0.476	1.000
Prune	81.0	0.1185	0.832	57.6	0.180	0.763	55.2	0.424	0.891
Compact (LLM)	82.1	0.126	0.884	70.7	0.222	0.941	58.2	0.444	0.933
Recency + Summary	82.1	0.117	0.821	71.2	0.179	0.758	58.5	0.435	0.914
Semantic Retrieval	78.0	0.120	0.842	67.1	0.162	0.686	55.3	0.419	0.880

Appendix

Table 8: Overall task performance and mean inference cost for DeepSeek V4 Pro. Performance is measured across all tasks; cost is measured on the same tasks that trigger context management when accumulated execution history exceeds

B=6

K. Cost is in reference USD; normalized cost uses uncapped Full history (

=1

) on the same tasks in each benchmark.

Agent runtime	Insertion point	Integration changes	Used for
opencode	experimental.chat.messages.transform plugin hook; the plugin is specified in opencode.json	The runtime’s own compaction and pruning are disabled so that only ContextRender manages historical context.	Integration only
AppWorld ReAct agent	trimmed_messages property replaced by a mixin on the live agent object	One property. Prompting, parsing, retry logic, tool use, runner, configuration, and CLI are unchanged.	AppWorld
ACON UnifiedAgent (smolagents)	MemoryManager subclass overriding get_conversation_history ; the class reference is rebound before agent construction	Memory class only. Runner, agent, prompts, retriever, and grading are unchanged.	8-objective QA
OpenClaw (CLI in Docker)	OpenAI-compatible proxy registered as a custom provider in openclaw.json	No runtime code is modified; context rendering is applied by the proxy before the request is forwarded.	Sanity check
Meta ARE default agent (GAIA2)	The same proxy, selected with the --provider local --endpoint flags	No runtime code is modified.	Integration check

Appendix

Table 9: Agent runtimes connected to ContextRender . For each runtime, we list the adapter insertion point, the runtime component modified by the integration, and its use in this work. Prompting, tool execution, parsing, and grading are not rewritten.

Explore similar work

FlowState: Execution State as Memory for Long-Horizon LLM Agents
Sep 28, 2026Minghao Li, Bangyan Li, Zifan Wang +5Long-Horizon Agents Long-Horizon Task Planning
Ant International, Ant Group
Plans Don't Persist: Why Context Management Is Load Bearing for LLM Agents
Jun 22, 2026Aman Mehta, Anupam DattaLarge Language Model Agents Long-Term Agent Memory
Snowflake AI Research
ContextWeaver: Selective and Dependency-Structured Memory Construction for LLM Agents
Apr 24, 2026Yating Wu, Yuhao Zhang, Sayan Ghosh +4Large Language Model Agents Contextual Reasoning
UT Austin · AWS AI Labs

Sep 28, 2026, cs.AIJ/K move · Enter open · S save

FlowState: Execution State as Memory for Long-Horizon LLM Agents

Minghao Li, Bangyan Li, Zifan Wang, Yulong Li+4

Ant International, Ant Group

Long-Horizon Agents Long-Horizon Task Planning Long-Term Agent Memory Persistent State Large Language Model Agents Persistence+2

Open paper AskPDF arXiv

Long-horizon tasks require LLM agents to continually draw on information from earlier interactions. However, retaining the full history increases context costs, while compressing it risks losing details needed later, and the relevance of historical information often becomes apparent as the task progresses. To address these challenges, we propose FlowState, which treats execution state as memory that can be retained and revisited across requests, unifying current decision-making with the reuse of historical information. FlowState preserves semantically typed state nodes, their relations, and references to raw tool observations, separating persistent retention from on-demand access. Within a single execution loop, Incremental State Update (ISU) maintains the current state based on new inputs and feedback, while Progressive State Access (PSA) progressively reveals historical states and supporting evidence as needed during reasoning. Together, these mechanisms enable agents to reassess prior decisions in light of new information and guide subsequent actions. Compared with a full-context baseline using the same DeepSeek-V4-Flash model, FlowState improves the average success rate on MemoryArena and the average pass rate on

τ^3

-Bench by 4.55 and 13.95 percentage points, respectively, while reducing total token consumption by 43.2% and 40.6%. These results demonstrate the performance and efficiency advantages of FlowState on long-horizon tasks.

Abstract

Explore similar work

FlowState: Execution State as Memory for Long-Horizon LLM Agents

Plans Don't Persist: Why Context Management Is Load Bearing for LLM Agents

ContextWeaver: Selective and Dependency-Structured Memory Construction for LLM Agents