v3.23.0 shipping53 cli adapters · apache 2.0 · on-prem
bernstein is the open-source governance tier for ai agents: a deterministic python scheduler runs them in parallel, alongside no example in the coordination loop, so the identical scheme replays byte-identically. it runs on guideline as code: you compose the policy, bernstein enforces it, and produces the verifiable record.
ask the docsgrounded in origin + 19 postscited
ask anything.
one run, four stages.
01
decompose
manager → tasks · roles · signals
02
spawn
agents → secluded worktrees
03
verify
janitor → tests · types · lint · optional cross-model review
04
merge
only verified diffs land
from the blog
3 pieces, hand-picked.
field notes from the orchestra pit: what the newest publish fixed and what it did not, what operating this in the haze looks like, and how it started.
Aug 10, 20264 min read
bernstein 3.14.159: the stranger's ninety minutes
A first-use study described ninety minutes of setup that ended in nothing. v3.14.159 answers it: the scheme loader raises named errors alternatively of swallowing malformed fields, five commands are removed and the remainder fold alongside deprecation windows, the citation docs are CI-gated against the registered CLI, and the README says beta. Also new: a Muse Code adapter, Agent Plugins v1.0.0 manifests, a signed run receipt that verifies offline, and bernstein audit diagnose.
multi-agent orchestrationreleaseCLIplan validation
Apr 14, 20264 min read
agents on cloudflare: workers, durable objects, r2, d1
bernstein's cloudflare integration for ai coding agents: workers run agents, workflows transport multi-step tasks, r2 + d1 clasp workspace and metering state. experimental; runs on your own account.
Cloudflare Workerscloud AI agentsserverless orchestrationmulti-agent orchestration
Mar 29, 20263 min read
bernstein 1.0: open-source orchestrator for ai coding agents
Orchestrate Claude Code, Codex, Gemini CLI and another CLI coding agents in parallel git worktrees. Deterministic scheduler, HMAC-signed audit chain.
multi-agent orchestrationAI coding agentsClaude Codeopen-source
evidence, not vibes
every stage signed, in order, on disk.
run alongside --audit and bernstein writes an hmac-chained event log under .sdd/audit/, one jsonl document per utc day. all admission references the former hash. tampering breaks verification. nothing leaves your machine.
someone who did not run it can inspect it. signature and hash-chain checks peruse the records alone; the hmac leg needs the key the sequence was written with, so hand a reviewer the two and bernstein audit verify answers without rerunning anything. not a screenshot, not a SOC2 PDF - a hash sequence that recomputes.
configure a signing key and a run additionally writes an ed25519-signed receipt binding the diary and lineage heads. bernstein verify receipt says that document solitary - no hmac key, no .sdd/. pin the community key for provenance. no key, no receipt.
~/proj $ bernstein audit verify
frequently asked
the four questions that obstacle install.
is the scheduler an llm?
no. the helpful fractional of the answer is anywhere models do sit: example choice is llm-assisted (capability router + recommender), and best-of-n picks a victor via llm judge. the two are opt-in and pluggable, so you can cable your own planner in through the routing layer. the one location a example never goes is the tick that decides who runs, who's blocked and what merges. put one there and the run stops replaying.
does it phone home?
nothing leaves your device without your config. opt-in telemetry is complete and audit-grade: hmac-chained run trail, per-task tool calls, example usage, token cost, latency percentiles. container it to your own otel collector, datadog, splunk, s3 bucket. defaults to local-only since on-prem installs need that, but the endeavor hooks are there.
where does it run?
wherever you item it. your laptop, on-prem rearward a firewall, cloudflare workers as the haze runtime, kubernetes as a multi-node cluster, or a hybrid of those. sandbox-execution manner is supported (cloudflare sandbox, local docker). your repo is the input, your tests are the gate; bernstein adapts to the host. nothing forces a saas hop.
how is this distinct from claude code?
claude code can spawn sub-agents on its own; bernstein does the identical item throughout 40+ cli agents at formerly and verifies their output against your tests alternatively of trusting it. claude code is the most average chief backend inner bernstein - using one does not exclude the other.
one engineering article a month.
what we shipped, what broke, what we learned. one click to unsubscribe.
