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/hutusi__80247856
userbench/penso__99f8e6bd
userbench/FSM1__e064c641
userbench/gabadi__566903d8
userbench/kungfusaini__76670d24
userbench/lyston11__581de3ef
userbench/yutakobayashidev__0341b973
userbench/nosman__75355bce
userbench/ET-NoahDolev__2951d3d1
userbench/ababushkin__e2b5880e
userbench/oddessentials__0b10a78b
userbench/lyston11__c12cd364
userbench/ET-NoahDolev__63f9b284
userbench/Stark-Industries0417__f054d028
userbench/christso__64756dde
userbench/hutusi__2751eb8c
userbench/dcambria__10d6ba09
userbench/gabadi__9e498963
userbench/blackgirlbytes__b2b09b06
userbench/dc_001__4b92ecaf

Displaying 20 of 620 tasks