Seeking entry-level Data Science / Software Engineering roles

RésuméContact
Go back

nowharness

Public

NowHarness

The interesting part of an agent is not the model, it is everything around it: tool selection, how much of the page the model reads, whether claimed work actually happened, and what stops unsafe writes. NowHarness builds those pieces one at a time and measures each against WorkArena. The evaluation layer came before the agent, a failure taxonomy decided what to build next, and every candidate feature was kept or reverted based on seed-matched re-runs. The adopted configuration raised task success from 53.3% to 83.3% on six held-out task types while cutting output tokens 61% and median latency 46%. Results live on a public dashboard.

+30 pts
task success (held-out)
-61%
output tokens
~860
logged episodes

The problem

Agent leaderboards mostly rank models, but enterprise reliability lives in the harness around the model: which tool gets selected, what the model has to read, whether the work it claims actually happened, and what stops a write it should not make. I wanted to build that layer piece by piece and prove, with statistics, which pieces earn their keep.

Approach

  1. 1.Built the evaluation layer before the agent, so every reported number traces to one of ~860 logged episodes. Two ground rules: an HTTP 200 is tool success, not task success, and for anything that writes, the model proposes and deterministic policy disposes.
  2. 2.Started with a deliberately basic agent, labeled its failures into a taxonomy, and let the taxonomy decide what to build next.
  3. 3.Ran ten experiment versions, changing one thing at a time with seed-matched re-runs. Two features were adopted: deterministic verification (28.6% cheaper at flat success) and structured tools (every previously-0% filter task taken off zero, Fisher p = 0.0016).
  4. 4.Rejected five features with their mechanisms explained, including a +6-point compression 'win' killed by seed-matched confirmation. The bar that killed the false win is the same bar the adopted features cleared.
  5. 5.Validated on six held-out task types that influenced zero engineering decisions, then published every run to a live dashboard.

Result

On the held-out suite the adopted harness raised task success from 53.3% ± 13.9 to 83.3% ± 11.8 while cutting output tokens 61% and median latency 46% (5 seeds per arm). On holdout, the agent discovered untaught ORDERBY sorting through the structured tool, going 0/5 to 5/5.

Stack

PythonPlaywrightWorkArena / BrowserGymSQLiteGitHub Pages