The brief history of Meta-Harnesses is a little undocumented, but it roughly goes: at first there was Conductor and Zed’s ACP, then there came OpenInspect, Cloudflare’s Flue, and then Vercel’s Eve and HarnessAgent, and Heypi.

It should not go unnoticed that today’s podcast guest Matei Zaharia, CTO of the enormously successful (for a pre LLM era company) Databricks, has a big bet now on meta-harnesses - Omnigent, an open source, pluggable architecture for pulling in any coding or knowledge work agent into a standardized, secure, reliable, scalable system:

It’s unclear whether or not Omnigent has the same kind of ingredients that made MCP’s success inevitable, but it is clear on an architectural level that some open source architecture that looks like this will probably win, if only because it is currently being independently rediscvoered at 1000 AI native shops.

AI News for 6/23/2026-6/24/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

OpenAI’s Jalapeño Chip and the Race Toward Full-Stack AI Infrastructure

• OpenAI goes deeper into hardware: OpenAI announced Jalapeño, its first custom AI chip for LLM inference, built with Broadcom and intended for ChatGPT, Codex, API traffic, and future agent products. The strategic message is straightforward: own more of the stack—chips, kernels, memory, networking, scheduling, deployment—so compute economics and product behavior become less dependent on merchant GPU supply. @gdb emphasized strong performance-per-watt, while @kimmonismus highlighted the reported 9-month design-to-tapeout cycle, unusually fast for a high-performance ASIC and reportedly accelerated by OpenAI’s own models.

• Technical read-through and ecosystem implications: Community reverse-engineering suggests Jalapeño looks TPU-like: @scaling01 estimated a near-reticle die, roughly 216GB HBM3E, ~7.1–7.4 TB/s bandwidth, and ~10 PFLOPS FP4. Even if those numbers remain unofficial, the signal is that hyperscaler-style inference silicon is now table stakes for frontier labs. The same day also reshaped the compiler/runtime landscape: Chris Lattner announced Qualcomm is acquiring Modular, while Modular said Mojo open-sourcing remains on track. That combination points to more serious competition around vertically integrated inference stacks beyond NVIDIA/CUDA.

• Serving and throughput remain active fronts: On the infra side, NVIDIA said NeMo AutoModel delivers 3.4–3.7x higher training throughput for MoE models via Expert Parallelism, DeepEP, and TransformerEngine kernels. SkyPilot launched Endpoints for unified inference across owned clusters, and Modal claimed open-source inference setups outperforming proprietary providers on latency. For local optimization, @jon_durbin reported 30–50% real-world decode gains from training custom DFLASH draft/speculator models.

Agent UX Shifts From “Tool” to “Coworker,” Raising New Security and Cost Questions

• Anthropic’s Slack-native agent model is the big UI story: Several tweets converged on the significance of Claude embedded into Slack/team workflows. @karpathy argued people are underrating it because it is not “just a feature” or Slack bot, but an org-level harness. @gallabytes described the experiential jump from Claude Code as a “pairing partner” to Tags as “managing a team.” @dabit3 pushed the idea further: eventually, you may not even need to explicitly tag agents.

• The hard part is identity, permissions, and lock-in: Anthropic detailed its agent identity model in this thread: Claude gets its own credentials, actions are auditable under that identity, and access can be revoked centrally. That design drew both praise and concern. @KentonVarda argued explicit per-agent permissioning does not scale and advocated capability-based security with fine-grained, task-scoped access. @random_walker framed Claude Tag as “a coworker that remembers everything and bills by the thought,” warning of tacit-knowledge lock-in, prompt-injection risk, and budget opacity once one shared agent becomes deeply embedded in org workflows. @JubbaOnJeans similarly flagged attribution ambiguity for write actions and future access-control complexity outside clean Slack-like boundaries.

• The open/DIY response is immediate: Hugging Face described its internal Slack-based coding agent Moon Bot in a blog tweet, emphasizing self-hosting, custom tools, auditable sessions, and zero lock-in. A follow-up from @calebfahlgren listed production integrations spanning GitHub, Athena, analytics, MongoDB, Elasticsearch, and HF Buckets. The larger pattern: teams increasingly want agent-native UX, but many would rather own the harness and memory layer than outsource organizational intelligence to a vendor.

Qwen-AgentWorld, OpenThoughts-Agent, and Memory as the Next Agent Scaling Axis

• Qwen-AgentWorld pushes “language world models” for agents: Alibaba Qwen introduced Qwen-AgentWorld, positioning it as a native language world model that simulates 7 environments—MCP, Search, Terminal, SWE, Web, OS, Android—inside a single model. Qwen claims two paths: build the simulator itself, and use world modeling as agent pretraining. They open-sourced Qwen-AgentWorld-35B-A3B and AgentWorldBench, with a 35B MoE / 3B active, 256K context model. One notable result: single-turn environment prediction transfers to multi-turn agent tasks with gains across both in-domain and out-of-domain benchmarks, as summarized in this follow-up.

• OpenThoughts-Agent contributes a serious open data recipe: @iScienceLuvr and @RichardZ412 highlighted OpenThoughts-Agent, an open curation/training pipeline for agentic models with 100+ controlled ablations. The team builds a 100K-example training set and fine-tunes Qwen3-32B, reaching 44.8% average accuracy across seven agentic benchmarks. The key findings are useful for practitioners: instruction choice matters disproportionately, strongest benchmark teacher ≠ best teacher, longer execution traces help, and source diversity beats over-repetition at scale.

• Memory is turning into a first-class systems layer: A lot of high-signal discussion centered on memory as the unresolved problem in agents. Weaviate’s Engram GA frames memory as asynchronous infrastructure that extracts, deduplicates, reconciles, and scopes memories rather than dumping everything into context. @hwchase17 showed a LangSmith/Context Hub workflow for “sleep-time compute,” where traces are analyzed offline and written back as memory. @dair_ai pointed to a paper arguing agent memory should be evaluated as a full data-management layer—storage, retrieval, update, consolidation, lifecycle—not a black box judged only by end-task success. This is increasingly where agent differentiation appears to be moving.

Chinese Open Models Keep Closing the Gap: GLM-5.2, Kimi Distribution, and Compute Scale

• GLM-5.2 continues to dominate the open-model conversation: Multiple tweets positioned GLM-5.2 as the strongest open-weight contender right now. CoreWeave said it tops open-model rankings on Artificial Analysis and Agent Arena, while Baseten and Cursor availability showed rapid serving/distribution uptake. @nutlope compared GLM 5.2 against Opus 4.8 on web tasks, reporting similar quality, ~2x token output, but still faster and roughly 3x cheaper. Arena also said GLM-5.2 Max leads Code Arena: Frontend against a strong field.

• Benchmark nuance matters: GLM-5.2 also showed up on ARC-AGI-2. @fchollet called it the strongest ARC-AGI-2 result to date by an open-source model, while others debated what its 22.8% really implies relative to frontier Western models. The broader takeaway is less about any single benchmark and more about open Chinese models being consistently “in the room” across coding, agents, and knowledge work.

• Commercialization and infrastructure acceleration: Moonshot’s Kimi API is now on AWS Marketplace, easing enterprise procurement via consolidated billing and EDP drawdown. Meanwhile, Chinese domestic compute remains a major theme: @teortaxesTex flagged reports that Huawei may demo a 950 SuperPOD scale system, implying production of large domestic NPU clusters at meaningful scale. If true, that would materially improve the economics and resilience of China’s model-serving ecosystem.

Policy, Talent, and Frontier-Lab Strategy Are Reshaping the Competitive Landscape

• Anthropic remains at the center of policy disputes: @kimmonismus reported the first major legal challenge to Trump-era AI export controls, with Legion arguing hosted model access is not equivalent to exporting weights or technical data. In parallel, the much-discussed Mythos story gained context: Reuters/AP details summarized here suggest Anthropic’s model found vulnerabilities in sensitive U.S. systems during a restricted testing exercise, though some commenters warned earlier coverage had been overstated.

• Distillation and access control are becoming geopolitical issues: @kimmonismus also reported Anthropic’s accusation that Alibaba-linked operators used ~25,000 fraudulent accounts and 28.8 million Claude exchanges to distill frontier capabilities into Qwen-class systems. If accurate, that escalates the “adversarial distillation” debate from rumor to something closer to enforcement and statecraft.

• Talent and new labs: The day also brought talent movement and new institutional formation. Arthur Conmy joining Anthropic is notable on the alignment side. Mirendil AI launched with a $200M seed round and a thesis around self-accelerating AI R&D for science. In the UK, BOLD Lab and SOFAIR received £60M in seed funding across two new national fundamental AI labs, with UCL DARK merging into BOLD. And on the commercial side, Bloomberg-reported departures from Google DeepMind toward Anthropic underscore how startup upside is continuing to pull frontier talent.

Top Tweets (by engagement)

• OpenAI Jalapeño: OpenAI announces its first custom inference chip — the most consequential product/infra launch in the set.

• GPT-5.5 Instant update: OpenAI rolls out a revised GPT-5.5 Instant with improved intent understanding, constraint handling, and conversational style.

• Qwen-AgentWorld: Alibaba Qwen launches and open-sources language world models for agents.

• Anthropic’s agent identity model: Claude in Slack now uses its own credentials and audit trail, clarifying one of the thorniest enterprise-agent design questions.

• Cursor x Notion: Cursor tasks can now be delegated directly from Notion, another sign that agent workflows are moving into existing team software rather than living in standalone chat apps.