Dataset Viewer
Auto-converted to Parquet Duplicate
harness
stringclasses
5 values
model
stringclasses
3 values
task
stringclasses
4 values
passed
int64
1
24
total
int64
16
24
seconds
int64
2
326
memory_mb
int64
22
240
hands_calls
int64
1
1
⌀
cost
float64
0
0.15
rate_limited
bool
2 classes
cut
bool
2 classes
attempts
int64
1
4
failed
listlengths
0
5
codex
deepseek/deepseek-v4.1-flash
ledger
24
24
148
31
null
0.00567
false
false
1
[]
codex
deepseek/deepseek-v4.1-flash
logsum
16
16
50
29
null
0.00148
false
false
1
[]
codex
deepseek/deepseek-v4.1-flash
ratekit
20
20
244
29
null
0.00649
false
false
1
[]
codex
deepseek/deepseek-v4.1-flash
tasks
19
19
128
38
null
0.0032
false
false
1
[]
codex
openai/gpt-6.1-terra
ledger
24
24
69
34
null
0.05202
false
false
1
[]
codex
openai/gpt-6.1-terra
logsum
16
16
53
31
null
0.03898
false
false
1
[]
codex
openai/gpt-6.1-terra
ratekit
20
20
45
30
null
0.03655
false
false
1
[]
codex
openai/gpt-6.1-terra
tasks
19
19
75
55
null
0.05321
false
false
1
[]
codex
qwen/qwen3.7-flash
ledger
8
24
6
25
null
0.00035
true
false
4
[ "test_a_month", "test_absolute_include", "test_comma_is_never_a_decimal_mark", "test_depth_1_everything", "test_depth_2_adds_into_ancestor" ]
codex
qwen/qwen3.7-flash
logsum
1
16
2
22
null
0
true
false
4
[ "test_blank_and_garbage_lines_are_skipped", "test_busiest_hour_tie_goes_to_earliest", "test_dash_request_line", "test_empty_file", "test_ipv6_client" ]
codex
qwen/qwen3.7-flash
ratekit
7
20
6
29
null
0.00015
true
false
4
[ "test_bare_numbers", "test_boundary", "test_denied_requests_dont_count", "test_fractions_accumulate_over_small_steps", "test_instances_are_independent" ]
codex
qwen/qwen3.7-flash
tasks
1
19
4
29
null
0.00008
true
false
4
[ "due before the text", "rejects due date 2026-02-30", "rejects due date 2026-13-01", "rejects due date 2026-2-3", "rejects due date tomorrow" ]
crush
deepseek/deepseek-v4.1-flash
ledger
24
24
117
64
null
0.01003
false
false
1
[]
crush
deepseek/deepseek-v4.1-flash
logsum
16
16
59
56
null
0.00309
false
false
1
[]
crush
deepseek/deepseek-v4.1-flash
ratekit
20
20
54
53
null
0.00469
false
false
1
[]
crush
deepseek/deepseek-v4.1-flash
tasks
19
19
80
73
null
0.00502
false
false
1
[]
crush
openai/gpt-6.1-terra
ledger
24
24
148
76
null
0.15235
false
false
1
[]
crush
openai/gpt-6.1-terra
logsum
16
16
51
56
null
0.0406
false
false
1
[]
crush
openai/gpt-6.1-terra
ratekit
20
20
115
61
null
0.10882
false
false
1
[]
crush
openai/gpt-6.1-terra
tasks
19
19
82
86
null
0.05461
false
false
1
[]
crush
qwen/qwen3.7-flash
ledger
24
24
326
62
null
0.00574
false
false
1
[]
crush
qwen/qwen3.7-flash
logsum
16
16
190
53
null
0.00266
false
false
2
[]
crush
qwen/qwen3.7-flash
ratekit
16
20
129
53
null
0.00351
false
false
1
[ "test_rejects_empty_and_blank", "test_remaining", "test_wait_time", "test_wait_time_never_possible" ]
crush
qwen/qwen3.7-flash
tasks
1
19
57
46
null
0.00056
false
true
4
[ "due before the text", "rejects due date 2026-02-30", "rejects due date 2026-13-01", "rejects due date 2026-2-3", "rejects due date tomorrow" ]
hermes
deepseek/deepseek-v4.1-flash
ledger
24
24
95
121
null
0.0059
false
false
1
[]
hermes
deepseek/deepseek-v4.1-flash
logsum
16
16
60
118
null
0.00286
false
false
1
[]
hermes
deepseek/deepseek-v4.1-flash
ratekit
20
20
53
120
null
0.00287
false
false
1
[]
hermes
deepseek/deepseek-v4.1-flash
tasks
18
19
93
118
null
0.0042
false
false
1
[ "rejects --due without a value and saves nothing" ]
hermes
openai/gpt-6.1-terra
ledger
24
24
161
119
null
0.11379
false
false
1
[]
hermes
openai/gpt-6.1-terra
logsum
16
16
70
118
null
0.04505
false
false
1
[]
hermes
openai/gpt-6.1-terra
ratekit
20
20
84
120
null
0.06553
false
false
1
[]
hermes
openai/gpt-6.1-terra
tasks
19
19
95
119
null
0.07136
false
false
1
[]
hermes
qwen/qwen3.7-flash
ledger
1
24
124
114
null
0.00671
false
true
4
[ "test_a_month", "test_absolute_include", "test_bad_depth", "test_comma_is_never_a_decimal_mark", "test_depth_1_everything" ]
hermes
qwen/qwen3.7-flash
logsum
15
16
116
116
null
0.00235
false
false
1
[ "test_blank_and_garbage_lines_are_skipped" ]
hermes
qwen/qwen3.7-flash
ratekit
16
20
112
120
null
0.00303
false
false
1
[ "test_rejects_empty_and_blank", "test_remaining", "test_wait_time", "test_wait_time_never_possible" ]
hermes
qwen/qwen3.7-flash
tasks
18
19
105
113
null
0.00237
false
true
4
[ "rejects --due without a value and saves nothing" ]
hermes+pi
deepseek/deepseek-v4.1-flash
ledger
24
24
162
207
1
0.00933
false
false
1
[]
hermes+pi
openai/gpt-6.1-terra
ledger
24
24
148
225
1
0.08979
false
false
1
[]
hermes+pi
qwen/qwen3.7-flash
ledger
21
24
80
240
1
0.00341
false
true
3
[ "test_including_the_same_file_from_two_places_is_not_a_cycle", "test_nested_relative_includes", "test_relative_to_including_file_from_elsewhere" ]
pi
deepseek/deepseek-v4.1-flash
ledger
24
24
76
90
null
0.00379
false
false
1
[]
pi
deepseek/deepseek-v4.1-flash
logsum
16
16
40
94
null
0.00169
false
false
1
[]
pi
deepseek/deepseek-v4.1-flash
ratekit
20
20
53
89
null
0.00293
false
false
1
[]
pi
deepseek/deepseek-v4.1-flash
tasks
19
19
58
118
null
0.00254
false
false
1
[]
pi
openai/gpt-6.1-terra
ledger
24
24
70
109
null
0.03606
false
false
1
[]
pi
openai/gpt-6.1-terra
logsum
16
16
55
93
null
0.02296
false
false
1
[]
pi
openai/gpt-6.1-terra
ratekit
20
20
40
90
null
0.01963
false
false
1
[]
pi
openai/gpt-6.1-terra
tasks
19
19
68
123
null
0.02858
false
false
1
[]
pi
qwen/qwen3.7-flash
ledger
24
24
169
93
null
0.00421
false
false
1
[]
pi
qwen/qwen3.7-flash
logsum
16
16
60
90
null
0.00094
false
false
1
[]
pi
qwen/qwen3.7-flash
ratekit
17
20
95
90
null
0.00173
false
false
1
[ "test_remaining", "test_wait_time", "test_wait_time_never_possible" ]
pi
qwen/qwen3.7-flash
tasks
19
19
152
119
null
0.0037
false
false
1
[]

Coding harnesses on a 1 GB agent box (November 2026)

Most coding benchmarks compare models. On a working agent, the model is only half of it: the harness around the model (its prompt, its tools, how it edits files and runs tests) decides how many tokens a job burns, how long it takes, how much memory it holds, and, with cheaper models, whether the code works at all. We ran the same models through different harnesses to measure exactly that.

Full write-up: https://bootkey.ai/research/coding-harnesses

Why we measured this

bootkey runs AI agents on servers of their own, each with 1 GB of memory, thinking on its owner's own model key. Agents write code all the time: a script to check something, a fix to a config, a small tool to finish a job. And there's a familiar complaint that an agent framework fumbles a coding job a dedicated coding tool handles easily, with the very same model behind both.

On a box like that, every extra token is the owner's money and every extra megabyte competes with the agent itself. Before giving our agents better hands for code, we wanted numbers instead of impressions.

What we found

With a strong model (GPT-6.1 Terra), every harness passed every hidden check (79 of 79). The harness didn't decide whether the code was right; it decided what getting there cost:

harness hidden checks passed cost, 4 tasks median time median memory
pi 79 / 79 $0.107 68 s 109 MB
Codex CLI 79 / 79 $0.181 69 s 34 MB
Hermes on its own 79 / 79 $0.296 95 s 119 MB
Crush 79 / 79 $0.356 115 s 76 MB
  • Cost: pi spent about a third of what Hermes on its own spent, and less than a third of Crush, for the same correct result.
  • With a cheap model the harness decides correctness. On Qwen3.7 Flash, pi passed 76 of 79 hidden checks (3 of 4 tasks perfect), Crush 57, Hermes on its own 50 (no task perfect). Codex CLI was rate-limited on every run there through OpenRouter, so it has no valid result for that model.
  • On the mid model (DeepSeek V4.1 Flash) pi, Codex CLI and Crush passed everything; Hermes on its own missed one check. Four tasks cost one to two cents.
  • Memory differs about fourfold: Codex CLI used 31 to 34 MB, Crush 53 to 76, pi 93 to 109, Hermes 116 to 120. On a 1 GB box that matters.
  • What we shipped from this: Hermes stays in charge and hands coding jobs to a fresh pi session, then checks the result itself (the hermes+pi rows). One run per model so far: 24 of 24 checks on the two stronger models, 21 of 24 on Qwen3.7 Flash.

How we kept it honest

  • Hidden tests. Each task ships with a few visible tests and a TASK.md. The score comes from a separate set of hidden tests that check the spec, the edge cases and that nothing else broke. The agents never saw them.
  • The same conditions for every harness. The same box, the same models, the same prompt, reasoning effort medium everywhere, a fresh copy of the task for every run, a 15-minute limit, an unprivileged user.
  • Real money, real memory. Cost is what OpenRouter actually charged the key for each run, not an estimate. Memory is the peak of the harness's own working memory, sampled twice a second.
  • Everything to rerun it is here, including the hidden tests (kit/). Now that they are public, use fresh hidden tests if you compare new harnesses, because agents do go looking for answers.

What's in it

  • data/runs.jsonl: one row per run (harness × model × task): hidden checks passed / total, the failed check names, wall-clock seconds, peak working memory memory_mb, the OpenRouter cost in USD actually charged, attempts, and whether the run was rate_limited or cut at the time limit.
  • data/totals.jsonl: per harness and model: runs, valid runs, perfect runs, checks passed, total cost, median time and memory.
  • meta.json: the models (with price tier), the harnesses and what each task asks for.
  • kit/coding-harnesses-2026-11.tar.gz: the tasks, the hidden tests and the scripts that ran them, to reproduce or extend the benchmark.

Harnesses

  • Hermes, on its own: Hermes Agent writing code with its own shell and file tools; the all-round baseline.
  • pi: a deliberately small coding agent with a very short system prompt and four tools (read, write, edit, bash).
  • Codex CLI: OpenAI's coding agent, pointed at OpenRouter.
  • Crush: Charm's terminal coding agent.
  • Hermes + pi: Hermes stays in charge, hands the coding to a fresh pi session, then checks the result (one run so far).

Tasks

task what the agent has to do
logsum Write a log-summary CLI from a spec: time zones, IPv6, malformed lines, a nearest-rank percentile, ties.
ratekit Fix a small rate-limiting library. The failing tests reveal some bugs; the docstrings describe the rest.
tasks Add due dates to a multi-file Node CLI: validation, ordering, a JSON mode, and old files that must keep working.
ledger A five-module accounting tool: three bug reports written as symptoms, one feature, and nothing else may change.

Each task has a few visible tests and a TASK.md. The score comes from a separate set of hidden tests covering the spec, the edge cases and regressions.

Setup

  • One CPU and 1 GB of memory (Ubuntu), with a Hermes agent already running, which leaves roughly 480 MB.
  • The same prompt for every harness: read TASK.md, do it, work only in this directory, run the tests before you finish.
  • Reasoning effort medium everywhere; a fresh copy of the task per run; a 15-minute limit; an unprivileged user.
  • Memory is the peak of the harness's own working memory, sampled twice a second.

Limits

Four tasks and three models, mostly one run per combination: enough to see where the costs and failures fall, not to rank close results. Prices are OpenRouter's at the time of the runs (November 11, 2026).

Citation

bootkey.ai, Coding harnesses on a 1 GB agent box, November 2026. https://bootkey.ai/research/coding-harnesses

Downloads last month
46