Who sets the standard — and who could take their place?
v1 · Operator curation (Ki) · Last updated
A field guide, not a world championship. Each SOTA pick is an editorial judgment within the stated scope, based on the sources reviewed. We have not rerun the benchmarks or established an overall winner.
Launch research seed — Ki’s individual sign-off is pending. Author-reported results and community examples are labeled. Scores and voting are not connected yet.
A–G reuse Measure rubric v0.2 exactly. H–J are supplemental lenses, not extra Measure points. Ki’s setup is a challenger, not a verified #1; its loadout is a seeded overview, not public benchmark evidence.
Persists workflow history and replays it after a worker restart instead of restarting the entire agent task.
The agent integration separates model/tool activities from durable coordination, with retries and long-running approval flows.
Limit: A reliability architecture pick, not a measured uptime winner. Retried side effects still need idempotency; it does not repair arbitrary infrastructure.
Makes persistent state, resumable execution and human intervention explicit parts of an agent graph.
Its Deep Agents ecosystem adds planning, subagents and filesystem tools on top of that durable runtime.
Limit: An orchestration pick, not proof of a 24/7 autonomous operator. Merge/deploy policy and real completion rates remain application responsibilities.
Exposes supported glasses’ displays, microphones, cameras and speakers through a common application interface.
Publishes its SDK and device support matrix, including Even Realities and Vuzix hardware, so integration claims can be checked.
Limit: A glasses interoperability pick, not a voice-latency or watch-battery winner. Hardware capabilities vary; no comparable all-day glasses + watch trial was found.
One self-hosted assistant connects to everyday channels including WhatsApp, Telegram and Slack with local state and memory.
Its public showcase documents personal workflows beyond coding; the implementation is available to inspect and adapt.
Limit: Examples are community reports, not audited daily usage. Channel count is not life-domain coverage, and health/finance autonomy is not established.
Combines OS-enforced filesystem and network boundaries; the engineering report measures 84% fewer permission prompts in internal use.
The web environment’s Git proxy checks repository and branch scope before attaching credentials outside the sandbox.
Limit: Fewer prompts is not an attack-blocking score. Correct configuration, secret redaction and adversarial testing remain necessary; no immunity is claimed.
Have better evidence? Register a loadout with a reproducible run, its date, model and budget. Onboarding currently saves a local draft; it does not submit a public challenge or cast a vote.