userbench/UserBench

UserBench eval: 620 next-message tasks across 62 developers. Composer 2.5 classifies gold and predicted messages into multi-label acts; reward is set Jaccard/IoU.

Published 7/21/2026 by

harbor run -d userbench/UserBench

UserBench

UserBench is a Harbor eval for how well an agent can stand in for a real software engineer mid coding-agent session.

At each held-out turn, the agent reads the conversation so far (/sim/history.md) and writes the single next developer message to /sim/answer.txt. A judge labels that message into a 4-way move taxonomy; reward is 1.0 iff the predicted move matches the gold move (move-match), else 0.0.

This package is the public noprofile cut: tasks ship history only. Developer style profiles are optional Harbor skills at job time (--skill / agents[].skills), not part of the task instruction.

Scale (this revision)

Developers 62 (everyone with ≥10 eval points; 6 with <10 dropped)
Tasks 620 (exactly 10 per developer)
Condition noprofile

Task selection rule

Same rule as build_agentic.py --per-dev N:

  1. Start from the held-out eval points in the clean UserBench v2 cohort.
  2. Keep developers with ≥10 points; drop the rest.
  3. For each remaining developer, sort points by environment/history.md byte size ascending (equivalent to sorting by len(context) in the builder) and take the first 10.
  4. Sessions may be shared across selected points for a developer. Ties break by legacy task directory name ascending.

Task naming

Harbor requires org/name package refs. Tasks are published as:

userbench/<username>__<hash>

Examples:

  • GitHub: userbench/winksaville__aad4a4f9
  • DataClaw (anonymized): userbench/dc_004__05fe03e7

Versioning is via Harbor tags only (e.g. @v2), not embedded in the task name.

Keywords

userbench · user-simulation · coding-agents · move-prediction · noprofile

Taxonomy / judging

Moves are one of: approve · critical · directive · inquiry (fault-first decision rule).

The task verifier classifies the agent’s predicted message (and gold real when gold_move is null) with a configurable judge (SIMBENCH_JUDGE, default Gemini). Published agentic numbers on userbench.vercel.app/results used Composer 2.5 as the job-time judge on a 226-point / 10-developer subset of the older package — those runs are not full-620 results.

Train / held-out split

Each developer has a deep train history and a strictly later held-out set. Eval tasks are prediction points carved from held-out sessions (noprofile: history for the held conversation only).

How to reference

What Ref
Dataset userbench/UserBench@v2 (also latest)
Task userbench/<username>__<hash>@v2
Hub https://hub.harborframework.com/datasets/userbench/UserBench
harbor run -d userbench/UserBench@v2 -a <agent> -m <model>
harbor run -p userbench/winksaville__aad4a4f9@v2 -a <agent> -m <model>

Links

Task
userbench/jakobtfaber__a0f851cc
userbench/ta93abe__14a96679
userbench/robouden__cb40a5e2
userbench/kungfusaini__6409200a
userbench/armelhbobdad__d8689cd9
userbench/khaong__06bbbed2
userbench/robouden__8b4b3c4a
userbench/thieso2__3dbff2d6
userbench/oddessentials__eaf6331c
userbench/johyunduk__4bab3ef7
userbench/Stark-Industries0417__9b04eb34
userbench/thieso2__67587edd
userbench/henryph24__82af8b05
userbench/jeevanpillay__d951d14a
userbench/armelhbobdad__248d4e03
userbench/kungfusaini__9a3333b8
userbench/jskswamy__2fd0350e
userbench/scottdensmore__7b129602
userbench/dcambria__31d94d0a
userbench/Soph__ed41ff61
userbench/TheurgicDuke771__fe9210ab
userbench/malkoG__f0e27113
userbench/jdsingh122918__f6d90727
userbench/kmiki0__924a7b32
userbench/dcambria__5bdc81bf
userbench/thieso2__5b029bc6
userbench/jhoetter__b49b7e7f
userbench/achildrenmile__7dafef0e
userbench/kungfusaini__f8d51bd4
userbench/ababushkin__0dafbf18
userbench/KeKs0r__a3482769
userbench/yyovil__149c4120
userbench/Poytr1__b2abde78
userbench/Poytr1__e42f2f97
userbench/gabadi__ae3008ee
userbench/raman325__d4a2df8b
userbench/manderson240__ec275534
userbench/gabadi__45c83dac
userbench/jdsingh122918__96fc2d3b
userbench/jhoetter__3517bfbd
userbench/melagiri__3cf814f5
userbench/thieso2__8f357020
userbench/jakobtfaber__0fddf0ed
userbench/singampalliveerendra__93f5702f
userbench/dc_000__4b92ddba
userbench/admarble__08f6ff97
userbench/jdsingh122918__f1bc9841
userbench/jhoetter__bcc3790b
userbench/blackgirlbytes__32cf9092
userbench/winksaville__98ba6868
userbench/cyyeh__93e85db5
userbench/alishakawaguchi__585afa4a
userbench/4thwithme__db958e1f
userbench/kmiki0__603c6c1f
userbench/malkoG__9b248afd
userbench/yutakobayashidev__f67ef680
userbench/wildlily1021__c031a215
userbench/dc_000__aa12035f
userbench/FSM1__8d0fb85e
userbench/yyovil__a5d654fb
userbench/Stark-Industries0417__69f46708
userbench/MohammedMqat__e972ce4f
userbench/jskswamy__e7cb0835
userbench/jakobtfaber__1395ade5
userbench/robouden__f12f7493
userbench/fcamblor__5f26916b
userbench/mvanhorn__716a17ba
userbench/dc_000__d5528693
userbench/scottdensmore__b162aa5c
userbench/jobinlawrance__a748127e
userbench/singampalliveerendra__13c48199
userbench/fcamblor__10653b23
userbench/admarble__1ffea695
userbench/fcamblor__41e0acf2
userbench/yutakobayashidev__23d09ae6
userbench/melagiri__7a34ab4a
userbench/junaid-appointy__9ae3d2e7
userbench/jobinlawrance__69d520ea
userbench/nathanbooth-konecta__273ef024
userbench/jobinlawrance__a1dc6ac0
userbench/jdsingh122918__235948db
userbench/khaong__7e2efac8
userbench/oddessentials__b4bcc213
userbench/lyston11__c145a945
userbench/wildlily1021__7748b2c8
userbench/henryph24__d2760f4a
userbench/Stark-Industries0417__8c531ba1
userbench/singampalliveerendra__8dc175b0
userbench/penso__d1d5889f
userbench/achildrenmile__210b9808
userbench/hutusi__490bc1e1
userbench/blackgirlbytes__077f6188
userbench/jobinlawrance__723f7ef8
userbench/Stark-Industries0417__fde8b674
userbench/barbogast__07017b77
userbench/raman325__9730850f
userbench/dc_004__633347c9
userbench/TheurgicDuke771__53dd0e94
userbench/gabadi__82b5fa6a
userbench/nosman__7c59d97b

Displaying 100 of 620 tasks