Datasets:

arsentev-ai
/

context-ucurve-coding-agents

run_id stringlengths 12 12	tasks_per_session int64 1 12	replicate int64 1 6	sessions int64 1 12	model_calls int64 58 107	input_tokens int64 116 214	output_tokens int64 37.3k 63.7k	cache_write_tokens int64 80.1k 265k	cache_read_tokens int64 2.95M 5.17M	total_tokens int64 3.13M 5.31M	modeled_cost_usd float64 2.02 2.98	wall_clock_s int64 409 690	tests_total int64 97 127	tests_passed int64 97 127
7ad4265d916f	1	1	12	95	190	41,780	257,116	3,667,239	3,966,325	2.6916	451	123	123
991ba46d5fe5	1	2	12	107	214	46,437	265,077	4,301,299	4,613,027	2.9816	667	117	117
b0ce6d78d965	1	3	12	95	190	51,081	257,084	3,639,806	3,948,161	2.8228	535	106	106
21bac64764c8	1	4	12	84	168	38,563	245,800	3,155,912	3,440,443	2.4475	466	114	114
c73de3aaae83	1	5	12	95	190	43,101	249,790	3,644,349	3,937,430	2.6771	465	122	122
7a919fd81c75	1	6	12	88	176	37,255	240,332	3,328,167	3,605,930	2.459	456	119	119
e95d7e6ab391	2	1	6	82	164	50,436	171,869	3,490,496	3,712,965	2.4487	519	117	117
efd7a83830d8	2	2	6	77	154	44,530	164,418	3,211,843	3,420,945	2.2485	522	105	105
17073b28a3db	2	3	6	85	170	48,174	167,413	3,605,145	3,820,902	2.4325	505	127	127
af42a745f337	2	4	6	95	190	43,781	174,534	4,065,676	4,284,181	2.5315	520	113	113
405173d7186b	2	5	6	79	158	39,566	168,367	3,344,213	3,552,304	2.2286	486	121	121
e38e518f05e9	2	6	6	88	176	49,037	171,421	3,754,937	3,975,571	2.5054	690	114	114
d5030af99b40	3	1	4	68	136	41,539	128,528	3,043,374	3,213,577	2.0185	409	120	120
3f1c0b6481b1	3	2	4	76	152	47,084	136,130	3,410,891	3,594,257	2.2405	469	111	111
a6280851367c	3	3	4	69	138	49,047	131,926	3,003,390	3,184,501	2.1319	521	114	114
20eeb24f2fad	3	4	4	76	152	51,955	134,441	3,346,568	3,533,116	2.2879	522	113	113
b70e95496c7c	3	5	4	72	144	44,515	129,576	3,203,916	3,378,151	2.1152	512	122	122
35ed101ee28b	3	6	4	68	136	45,155	126,002	2,954,962	3,126,255	2.0367	462	107	107
04fb08b01d56	4	1	3	69	138	52,604	122,286	3,375,927	3,550,955	2.2608	514	110	110
a314a80b451d	4	2	3	66	132	47,818	118,045	3,143,415	3,309,410	2.1034	458	122	122
c75ba3c5a591	4	3	3	74	148	53,960	132,063	3,708,824	3,894,995	2.4177	637	115	115
ba3fc35a6d31	4	4	3	74	148	57,404	129,621	3,674,635	3,861,808	2.45	570	112	112
0fc5b487c216	4	5	3	73	146	53,214	124,269	3,583,088	3,760,717	2.3396	520	124	124
3e0ef41dcb9d	4	6	3	70	140	45,034	114,584	3,314,962	3,474,720	2.1001	458	116	116
fe6dbd0555ed	6	1	2	66	132	46,805	97,438	3,310,994	3,455,369	2.0612	490	109	109
9ad717907bf6	6	2	2	63	126	46,802	99,882	3,225,309	3,372,119	2.0446	487	115	115
48dd7bcb4a33	6	3	2	77	154	52,880	110,047	4,245,681	4,408,762	2.48	630	100	100
bfc047f10a8c	6	4	2	70	140	50,159	103,416	3,691,313	3,845,028	2.248	504	113	113
84e41e3aa677	6	5	2	72	144	46,412	103,820	3,840,881	3,991,257	2.2382	452	106	106
b8d8da6086f3	6	6	2	78	156	54,813	119,872	4,467,254	4,642,095	2.6124	582	118	118
0be049a6acbd	12	1	1	60	120	49,072	82,563	4,097,568	4,229,323	2.2753	473	107	107
2e82091b7a73	12	2	1	63	126	55,255	89,944	4,506,184	4,651,509	2.5183	527	115	115
3b9e1f50905b	12	3	1	75	150	52,646	89,065	5,167,386	5,309,247	2.6743	501	105	105
c6f35f1bafb5	12	4	1	58	116	49,307	80,117	3,886,467	4,016,007	2.2063	526	97	97
df5656c84d8b	12	5	1	63	126	53,033	86,489	4,510,093	4,649,741	2.4732	581	114	114
dd698a8f51e6	12	6	1	59	118	63,705	98,898	4,540,502	4,703,223	2.6889	650	103	103

Context U-curve: 36 coding-agent runs under six context-clearing policies

How often should an LLM coding agent's context be cleared? This dataset holds every run behind the report "Clear Every Third Task: A Measured U-Curve in the Context Economy of Coding Agents" (Evgenii Arsentev, 2026, DOI 10.5281/zenodo.22699668).

A fixed suite of twelve programming tasks was run under six session-length policies — a fresh session every 1, 2, 3, 4, 6 and 12 tasks — with six replicates each, holding the model, the tasks, their order and the verification suite constant. 36 runs, 4086 tests, 0 failures.

Summary by policy

tasks per session	runs	mean modeled cost, USD	mean cache writes	mean cache reads	tests passed
1	6	2.68	253k	3.62M	701/701
2	6	2.40	170k	3.58M	697/697
3	6	2.14	131k	3.16M	687/687
4	6	2.28	123k	3.47M	699/699
6	6	2.28	106k	3.80M	661/661
12	6	2.47	88k	4.45M	641/641

Cost is not monotone in session length: clearing after every task and never clearing are both more expensive than clearing every three to six tasks. Cache writes fall with longer sessions while context carried per call rises; their product, cache reads, is U-shaped. Full statistics (exact permutation tests) are in the report.

Fields

One row per run. Counters only — no prompts, no model output, no file paths, no project names.

field	meaning
run_id	hash of the run directory name
tasks_per_session	policy: a fresh session every N tasks (12 = never cleared)
replicate	replicate number 1–6
sessions	sessions the run used
model_calls	API calls made by the agent
input_tokens / output_tokens	uncached input and model output tokens
cache_write_tokens / cache_read_tokens	prompt-cache writes and reads
total_tokens	sum of the four token counters
modeled_cost_usd	cost at the model's public list prices
wall_clock_s	wall-clock duration of the run
tests_total / tests_passed / tests_failed	verification suite result after the run

Tooling

The run-efficiency metric from this study is implemented in contextburn (pip install contextburn, MCP server ai.arsentev/contextburn).

Citation

@techreport{arsentev2026ucurve,
  author = {Arsentev, Evgenii},
  title  = {Clear Every Third Task: A Measured U-Curve in the Context Economy of Coding Agents},
  year   = {2026},
  institution = {ARSENTEV.AI},
  doi    = {10.5281/zenodo.22699668}
}

Author: Evgenii Arsentev · ORCID 0000-0002-9120-7298 · arsentev.ai/research

Mirrors and related records

Kaggle copy of this dataset: https://www.kaggle.com/datasets/arsentevai/context-u-curve-of-coding-agents-36-runs

OSF project with the report PDF and the data: https://osf.io/5qtwy/ (DOI 10.17605/OSF.IO/5QTWY)

Demo of contextburn over these runs: DOI 10.5281/zenodo.22713920, https://www.youtube.com/watch?v=ep7LXFernwQ, https://archive.org/details/contextburn-demo-context-ucurve-2026

Podcast episode about this measurement: https://arsentev.ai/podcast/when-to-clear-agent-context

All reports, datasets and DOIs: https://arsentev.ai/research

Downloads last month: -

Number of rows:

Total file size:

9.19 kB