live
GATE
trace-vault
Snapshot testing & replay for AI agents — CI for non-deterministic systems.
Core insight: determinism ≠ faithfulness. A reproducible run still has to prove it did the right thing.
Agent engineer · verification systemsAgent 工程师 · 验证系统
I build the trust layer for AI agents. 我为 AI agents 构建会自证的信任层。
The demo is not the finish line. I build replay harnesses, black-box ledgers, eval gates, and proof receipts that make autonomous agents dependable after the human stops watching. Demo 跑通不是终点。我做 replay harness、黑匣子账本、eval gate 和 proof receipt，让自主 agent 在没人盯着时也能被验证、被追责、被信任。
agent.run
attested
trace
eval
receipt
proof
PASS
Scroll-driven verification滚动驱动的验证叙事
Trust is a sequence, not a slogan.信任是一条链路, 不是一句口号。
The page moves like the system I build: a run starts as a claim, passes through independent records and replay, then exits as a signed verdict. 这个页面的滚动方式就是我的产品观: 一次运行先只是声明, 经过独立记录与重放验证, 最后才变成可签收的裁决。
-
01
Demo worksThe agent completes a real task.
-
02
Trace recordedEvery event lands in an append-only ledger.
-
03
Dual witnessThe agent's receipt meets an independent black box.
-
04
Eval gate passedClaims are replayed instead of trusted.
-
05
Proof receipt signedOne verdict format ships to CI.
trust-report
$ trace-vault replay agent-run.json
· reading ledger ........ idle
· witness cross-check ... idle
· eval gate ............. idle
· proof receipt ......... pendingsha2560000-0000
One ecosystem一个生态
Not isolated projects. One agent trust stack.不是散落项目, 是一套 agent trust stack。
RUN
Alfred
machine-gated autonomy, signed receipts
RECORD
NightWatch
independent black box, claims re-verified
GATE
trace-vault
offline replay, determinism ≠ faithfulness
Two contracts make them one system: Claude Code-compatible hooks, and one Agent Trust Report v0 verdict that CI can read. 两个契约把它们焊成一个系统: Claude Code 兼容 hooks, 以及 CI 可读取的同一种 Agent Trust Report v0 裁决格式。
Specimen cards项目标本
Systems that leave evidence.会留下证据的系统。
npm
RUN
Alfred
A verifiable autonomous coding agent in your terminal.
“Done” is a machine-enforced verify gate; every hands-off run leaves a signed, replayable ledger.
site
RECORD
nightwatch
The black box recorder for overnight AI agents.
Every session event lands in a hash-chained append-only ledger, with morning debriefs that verify the agent’s claims.
wip
PROVE
provenant
Glass-box bill auditing with HMAC-signed proof receipts.
A provenance lens for cost, usage, and claims — built around receipts that can be checked without trusting the narrator.
About关于
Hi, I'm Beamus.你好, 我是 Beamus。
I'm an agent engineer focused on what happens after the demo works: making AI agents reliable, verifiable, and trustworthy enough to run without a human watching.
My bet for the loop-engineering era: when models write the code, the engineering that still matters is the loop around them — memory, replay, eval, provenance, and receipts.
Open to Agent Engineer & AI Engineer roles
Writing & guides写作与教程
Notes from the loop.来自循环里的笔记。
Agent Trust Report v0 — the spec
One verdict format three gates emit, plus a real dual-witness run with raw ledgers committed.
SPEC ↗
RAG, from zero to production
A Chinese learning path: LangChain basics → GraphRAG → Agentic RAG → enterprise.
GUIDE ↗
Mastra for TypeScript developers
Build the mental model first, then Memory, RAG, MCP, eval, and deploy.
ZH ↗