UseAgent

Dogfood closeout · OSBlog

From a real workload
to better control.

UseAgent supervised a bilingual open-source blog through planning, worker fallbacks, audits, quota failure, takeover, QA and Vercel release evidence. This page keeps the successes and the friction visible.

Diagram of the shared repository ledger used to preserve OSBlog tasks, reports, evidence and checkpoints.
The repository became the durable memory for a real multi-agent release, not just a demo.
Production boundary

OSBlog's live target for this case study is Vercel: osblog.thuanlyt.id.vn and osblog.vercel.app. VPS, Netlify and local Node are documented targets, not live OSBlog environments.

What was actually evidenced

01Planning and dispatch

A light goal and roster became scoped work items, mailboxes and copyable worker prompts.

02Failure and recovery

Unavailable runtimes, quota exhaustion, no-report attempts and takeover history remained visible instead of being rewritten as success.

03Audit to fix

Independent review reproduced defects; workers added fixes and regressions; a second review checked the result.

04Live release boundary

Vercel aliases, SSR/SEO routes, feeds, sitemap, robots and existing Cap media have recorded smoke evidence.

Timeline

  1. Bootstrap.

    UA-0001 and UA-0002 established the goal, roster, ledger and first dispatch.

  2. Fallback.

    Unavailable Antigravity/Claude paths and failed Codex waits were preserved; supervisor-local recovery produced a labeled replay.

  3. Review.

    UA-0048 found four real application defects; UA-0052 fixed them and UA-0058 independently re-reviewed the runtime.

  4. Release.

    Vercel and recovery evidence accumulated, while the positive historical-308 fixture stayed intentionally uncreated.

  5. Takeover.

    UA-0091 exhausted quota; UA-0093 resumed from a partial checkpoint and passed its focused security gate.

Findings that changed the product conversation

The strongest result was not task count. It was the ability to distinguish live, local, simulation and blocked evidence.

Shipped: evidence and recovery

UA-0048 through UA-0051 now provide report freshness, typed provenance, takeover lineage and bounded sanitized evidence with raw diagnostics kept in a local spool.

Shipped: source and runtime gates

UA-0052 through UA-0055 bind QA and release durability to source state, default QA to safe argv execution, and worker intake to bounded readiness and failure classification.

Residual: higher assurance

Capture verification, stale-lock recovery, filesystem diff enforcement and authenticated remote identity remain backlog. They are not blockers for the documented trusted-local UseAgent model.

Capture manifest

The existing OSBlog Cap GIF and MP4 are committed, hashed and documented as a real public-site capture. The fresh UA-0086 capture was not promoted because its Brave window could not be verified. See the full capture manifest and the source-anchored Markdown case study.

Tóm tắt bằng tiếng Việt

OSBlog là workload thật để kiểm chứng UseAgent. Supervisor đã tạo task/DAG, dispatch cho Codex, Claude và Antigravity-style identities, giữ failure/quota history, đọc report, audit, QA, takeover và evidence Vercel. Media Cap cũ được giữ vì đã có provenance; capture mới bị block vì không xác minh được browser target.

Finding quan trọng nhất là repository evidence phải có provenance và source identity rõ. UseAgent hiện đã ship report freshness, typed evidence, takeover lineage, evidence hygiene, source-bound QA, Git durability, safe QA và runtime readiness; sandbox, identity xác thực và cycle metrics vẫn là backlog.

Read the full evidence record