Andrej Karpathy described “Claws” as a layer above LLM agents that adds orchestration, scheduling, context management, tool calls, and persistence; he was experimenting with the concept on a Mac Mini while expressing caution about OpenClaw specifically. Simon Willison highlighted the framing and suggested Karpathy may have found useful terminology for this emerging system pattern. The supplied results establish the label and architecture-level idea, but claims that it is already gaining enterprise traction—or will catalyze distinct implementations rather than rename existing assistants—come mainly from promotional or secondary sources and are not yet demonstrated.
Karpathy’s orchestration-and-persistence framing independently converges with Scott’s already-developed persistent-agent architecture and active OpenClaw implementation, creating a strong dated-receipts and positioning opportunity. The further hypothesis that the “Claws” label will catalyze distinct implementations remains unproven, but adoption of this framing could directly affect how Scott presents and differentiates OpenClaw.
ip:source.personal-agents-three-jobs-ebookip:framework.long-running-agentsip:framework.delegation-planedev:project.openclawdev:technology.openclawradar:person.andrej-karpathyradar:concept.agent-memoryradar:concept.agent-harnessesradar:karpathy-llm-wiki-adoption
queries asked of Scott's wikis
- persistent personal agent architecture
- agent harness orchestration scheduling context and tools
- always-on local agents sandbox security
- agent memory identity and long-running state
- messaging interfaces for proactive personal AI
- terminology catalyzing new software patterns
2026-07-26T16:22:49Z
Months after the original framing, persistent-agent implementations are clearly established but remain largely independent of Karpathy’s terminology, with no credible attribution showing that “Claws” catalyzed them. Further generic implementations are unlikely to resolve the causal claim, so the episode has exhausted its useful observation window.
2026-07-26T15:21:26Z
The refreshed discussion is low-engagement repetition around persistent-agent implementations and adds no attribution to Karpathy’s framing. The underlying category is established, but the case’s specific catalysis claim remains unsupported and should wait for explicit adoption or influence evidence.
2026-07-24T08:23:54Z
The 44-agent system adds implementation density to an already-established persistent-agent pattern but provides no evidence that Karpathy’s “Claws” framing influenced its design or terminology. The central catalysis claim remains open but increasingly detached from the repetitive implementation stream.
2026-07-24T08:21:10Z
evidence attached: reddit.post.1v54sb5 — A concrete 44-agent personal automation system is practical evidence that persistent personal-agent architectures are being implemented, albeit as a single anecdote.
2026-07-23T16:26:11Z
The latest activity adds no probative evidence that Karpathy’s framing catalyzed implementations; it remains repetitive amplification of an already-established persistent-agent pattern. Keep the causal claim open, but cool the case until an implementation explicitly attributes its design or terminology to the framing.
2026-07-23T15:24:04Z
Atlas adds another concrete persistent-memory, multi-agent implementation, further establishing the underlying category but again showing no adoption of—or attributed influence from—Karpathy’s “Claws” framing. The implementation evidence is now repetitive rather than advancing the case’s central catalysis claim.
2026-07-23T15:21:40Z
evidence attached: reddit.post.1v4gu3k — A concrete open-source implementation combines persistent local agent memory with a multi-agent coding fleet, bearing on whether persistent personal-agent patterns are taking shape.
2026-07-23T02:24:37Z
OmniDesk adds another concrete persistent-agent workflow, but it neither adopts the “Claws” framing nor shows influence from Karpathy. The broader pattern is now repeatedly demonstrated, while the case’s causal catalysis claim remains unsupported.
2026-07-23T02:21:00Z
evidence attached: reddit.post.1v3zt4s — A concrete implementation of agents finding, maintaining, and shipping work provides weak independent evidence that persistent personal agents are becoming a recognizable pattern.
2026-07-22T21:24:58Z
ClawLite adds a second implementation adopting “Claw” terminology, suggesting the label is becoming recognizable beyond Karpathy’s essay. With only a title, negligible engagement, and no stated influence from Karpathy, it still does not establish that his framing catalyzed the implementation rather than naming an already-emerging pattern.
2026-07-22T21:21:05Z
evidence attached: hn.story.49013033 — ClawLite is an independent implementation bearing on whether Karpathy's persistent-agent framing catalyzes real local-first assistants.
2026-07-22T14:30:08Z
The two HN implementations strengthen the broader persistent-agent category with event-driven waking and persistent cloud execution, but neither invokes the “Claws” label or indicates influence from Karpathy. The evidence is increasingly corroborative of the underlying pattern while remaining non-probative—and now somewhat repetitive—on the central catalysis claim.
2026-07-22T14:21:39Z
evidence attached: hn.story.49007062 — This is independent evidence that persistent cloud execution is becoming a concrete implementation pattern for personal and coding agents.
2026-07-22T14:21:39Z
evidence attached: hn.story.49006433 — Independent implementation corroborates the emerging persistent-agent pattern by waking Codex threads only when external events occur.
2026-07-22T07:25:55Z
The additional local workflow is another recognizable persistent-agent implementation, but it neither uses the “Claws” framing nor links its design to Karpathy. Evidence is accumulating for the underlying pattern while remaining repetitive and non-probative on the case’s central catalysis claim.
2026-07-22T07:21:02Z
evidence attached: reddit.post.1v38j97 — A concrete local implementation of queued work, persistent sessions, and agent-to-agent tickets supports the emerging persistent-agent pattern.
2026-07-21T19:28:53Z
The Claire Vo example is the first attached implementation explicitly using the OpenClaw/“Claws” language, narrowing the gap between the framing and deployed systems. However, it comes through a weak secondary Reddit account and does not establish that Karpathy’s framing caused the implementation, so the central catalysis claim remains uncorroborated.
2026-07-21T19:21:42Z
evidence attached: reddit.post.1v2qlal — A concrete production example of a solo founder operating multiple persistent, role-specialized agents materially supports the Claws implementation hypothesis.
2026-07-21T18:26:41Z
A second independent implementation strengthens the view that persistent-agent patterns are emerging, but neither implementation references or appears catalyzed by the “Claws” framing. The new evidence therefore reinforces the underlying category without advancing the case’s central causal claim.
2026-07-21T18:21:37Z
evidence attached: reddit.post.1v2ozg2 — This is a concrete implementation of persistent, multi-agent task coordination with human handoff, though its low engagement limits evidentiary weight.
2026-07-20T17:30:31Z
Oak shows that recognizable persistent personal-agent implementations are emerging beyond the original discussion, moving the case past a pure terminology seed. It does not reference Karpathy’s “Claws” framing, so the key claim that the label is catalyzing implementations remains uncorroborated.
2026-07-20T17:21:29Z
evidence attached: reddit.post.1v1rkpb — This is independent corroboration of a persistent personal agent that maintains structured user state and workflows over time.
2026-07-20T06:48:55Z
canonical anchor promoted hn.story.47099160 -> echo.x.37ccc8945a: the true origin is Karpathy's X mini-essay (quoted in full via Willison's blog). The tweet names OpenClaw directly. Echo latency 3582.0h — the story predates the radar by months.
2026-07-20T06:29:48Z
grounded: converges/high — Karpathy’s orchestration-and-persistence framing independently converges with Scott’s already-developed persistent-agent architecture and active OpenClaw implem
2026-07-20T06:24:05Z
grounded: known/medium — Scott already holds and implements the underlying position in “Long-Running Agents,” “Personal Agents Three Jobs,” and the active OpenClaw stack: persistence, s
2026-07-20T06:21:37Z
case created — A newly articulated pattern for persistent personal agents is directly relevant to agent-harness design but has not yet shown independent implementation.