Skip to main content

Compare

Bernstein vs ReferenceComputerUse: quick decision guide

Browser / computer-use adapter family (#2606).

Page built on 2026-08-10 from data/adapters-meta.json. Every claim below links to its primary source.

Install both

ReferenceComputerUse

No install command recorded in the bernstein adapter source as of 2026-08-10.

Bernstein

pipx install bernstein

Apache-2.0. Deterministic Python scheduler.

Feature matrix

CapabilityReferenceComputerUseBernstein
Install methodNot recordedpipx install bernstein
LicenseNot recordedApache-2.0
AuthenticationNot recordedPer-agent credential scoping (no shared key)
Multi-agent orchestrationOne agent in a terminalReferenceComputerUse plus 40+ other adapters in parallel worktrees
MCP supportNot measuredYes
Parallel-safe in worktreesNot measuredYes (designed around git worktrees)
HMAC-chained audit logNoYes (RFC 2104 SHA-256 chain in .sdd/)
Deterministic schedulerNot applicable (single-agent CLI)Yes (Deterministic Python scheduler)

Adapter source: src/bernstein/adapters/computer_use.py

Verifiable facts

The brief for this surface requires at least three facts that a reader can verify against a primary source. The list below is built from the bernstein adapter source and, when available, the upstream project's own pages.

  1. Bernstein ships a ReferenceComputerUse adapter at src/bernstein/adapters/computer_use.py that wraps the upstream CLI as one of 42 routable agents. [source: bernstein adapter source, as of 2026-08-10]
  2. Bernstein is an open-source Multi-agent orchestrator licensed Apache-2.0, with a deterministic Python scheduler that routes work across CLI agents in parallel git worktrees. [source: bernstein repo, as of 2026-08-10]
  3. Bernstein records lineage and a replay journal on every run; with the audit log enabled it also writes an HMAC-SHA256 chained record under .sdd/ that lets a reviewer replay every routing and quality-gate decision. [source: bernstein repo, as of 2026-08-10]

Where ReferenceComputerUse fits in Bernstein

Bernstein registers ReferenceComputerUse under the slug "computer-use" and the registry name "computer_use". The adapter source lives at src/bernstein/adapters/computer_use.py in the bernstein repo and was last touched at build time 2026-08-10. The ReferenceComputerUse adapter file is 261 lines and 9,690 bytes long, fingerprinted 9b918a0d2ca3333a (first 16 hex chars of SHA-256). No upstream GitHub repository is recorded in the bernstein adapter for ReferenceComputerUse; refer to the upstream vendor's documentation when auditing. The bernstein adapter file for ReferenceComputerUse does not yet carry a "Last verified against upstream" line; this means the adapter still tracks an unpinned upstream binary. Bernstein routes tasks to ReferenceComputerUse when its pass rate on similar work clears the configured threshold, otherwise the deterministic Python scheduler picks a different adapter from the 40+ adapter catalog.

Adapter source excerpt

The module docstring of the ReferenceComputerUse adapter, as it stands in the bernstein repo. Punctuation is normalised for the web; the wording is the author's. Length: 1182 characters.

Browser / computer-use adapter family (#2606). This admits the one agent kind the coding-adapter substrate could not front: *third-party autonomous* browser and computer-use agents. A coding adapter's output is a file diff; a browser / computer-use agent owns its own decision loop and emits a stream of GUI actions. This adapter fronts such an external agent and leaves the per-action anchoring to :mod:`bernstein.core.agents.computer_use_attestation`, so the boundary never imports a specific browser tool -- the concrete driver stays behind the adapter contract. Two things live here: * :class:`ComputerUseAdapter` -- a :class:`~bernstein.adapters.base.CLIAdapter` subclass adding per-task profile isolation and the capability-refusal path, the same on-ramp the coding adapters use. * :class:`ReferenceComputerUseAdapter` -- a concrete reference adapter (registry name ``computer_use``) that proves the mechanism end to end without requiring a live browser in tests. Driver failure and timeout never surface as free text: they map onto :class:`ComputerUseTerminalState` via :func:`classify_terminal_state`, and a driver fault is raised as a typed :class:`ComputerUseDriverError`.

Adapter telemetry

Registry namecomputer_use
Adapter classReferenceComputerUse
Source filesrc/bernstein/adapters/computer_use.py
Source file size261 lines, 9,690 bytes
Source SHA-2569b918a0d2ca3333a84106536a284d6faa28d9aa96a0c0449be0567b8ea996f3e
Category bucketcli-family
Upstream repoNot derivable from adapter source
Upstream homepageNot recorded
Last verified upstreamNo "Last verified" line in adapter source
Operator-curated overlayNo (programmatic page)

When to pick which

Choose ReferenceComputerUse

Reach for ReferenceComputerUse when the work is a single thread that fits one agent: in a single-process terminal session, designed for single-instance use per repo. Auth model is configured per upstream docs. You skip the orchestrator round-trip and get the smallest possible surface between you and the model.

Choose Bernstein

Wrap ReferenceComputerUse under Bernstein when the goal splits into parallel tasks, when you want an HMAC-chained audit log on every routing decision, or when a deterministic Python scheduler (no LLM picking who runs what) is a hard requirement.

FAQ

Does Bernstein replace ReferenceComputerUse?

No. Bernstein wraps ReferenceComputerUse as one of 40+ CLI adapters and routes tasks to it based on per-task pass-rate history. ReferenceComputerUse keeps running unchanged; Bernstein decides when it gets work.

Can I run ReferenceComputerUse alongside other agents in the same repo?

Yes. Each agent runs in its own git worktree under .sdd/worktrees/{session_id}, so file edits never collide. Bernstein merges results back to the trunk only after the configured quality gates (lint, types, tests) pass.

Is this comparison page handwritten?

No. The template is fixed; every fact and every link is pulled from the bernstein adapter source in the master branch and (when available) the upstream project's own pages. The data extractor lives at scripts/gen-compare-data.mjs. No LLM writes the prose.