The Layer Above the Harness, and the Floor Beneath It
The operational layer around agents got a lot of attention this week — meta-harnesses, persistent cloud execution, engineering-org rules calibrated for autonomous execution. Read together, they describe a stack that is finally being built out above the model. Read against Anthropic's Friday-night Fable/Mythos suspension, they describe a stack whose floor can drop away on a phone call.
Will Larson's post is the most concrete artifact of what running at the upper layer actually looks like: an entire year of migrations, deploy pipelines, and process redesign predicated on the assumption that the base case of most work is now agentic. Databricks and OpenAI both shipped things this week that would make that easier for teams operating one tier lower — a common interface across Claude Code, Codex, Cursor, and Pi; a persistent cloud execution substrate for Codex sessions that outlive a laptop lid. The bet is that harnesses will multiply and models will churn, but the layer above them stays.
The Fable/Mythos suspension is the argument against that bet in its purest form. Artificial Analysis's Intelligence Frontier chart moved backward for the first time. Nothing in Omnigent's YAML or Codex's persistent sandbox reroutes around an export-control directive that revokes access overnight for a vendor sitting on top of the coding leaderboard. The meta-harness answer to model risk — swap providers with a one-line change — is a real answer, but it presupposes you have a competitor at parity to swap to. This week that assumption got weaker, not stronger.
Two stories collided this week, and together they reframe how to think about betting a roadmap on AI coding agents. The optimistic one: the tooling layer above the models is finally maturing. Databricks open-sourced a common interface that lets teams route work across Claude Code, Codex, Cursor and other agents interchangeably. OpenAI is buying its way into persistent cloud execution for Codex, so agent sessions can run for hours without a developer babysitting a laptop. And a widely-read engineering leader published a concrete account of an org that went from about six deploys a week to several hundred by making agentic work the default — a real data point for anyone still debating whether this changes engineering throughput.
The pessimistic one: on Friday night, a US export-control action suspended Anthropic's top coding models, and an independent frontier-capability tracker registered its first-ever backward move. The whole premise of the tooling-layer bet — that models are interchangeable and you can swap vendors with a config change — only works if a comparable competitor exists to swap to. This week that got harder, not easier. For procurement and build-vs-buy decisions, the implication is concrete: multi-vendor abstraction is now table stakes, but so is honest scenario planning for a primary vendor going dark on short notice. Concentration risk in this stack is a policy risk, not just a commercial one.
Two stories this week point in opposite directions. The first: companies are getting serious about building AI "agents" — systems that don't just answer questions but actually go off and do work, like writing and shipping code on their own. One startup went from releasing software updates a handful of times a week to hundreds of times a week after handing most of the coding over to AI assistants. OpenAI is buying a company to let its coding AI keep working in the background even after you close your laptop, and it now has around five million people using it every week. The tools that sit on top of these AIs are starting to look like real infrastructure.
The counter-story arrived Friday night, when the US government told Anthropic (the maker of Claude) to cut off access to one of its top coding models for certain customers. Overnight, a leading tool disappeared from the market for those users. For the first time, an industry tracker showed the frontier of AI capability moving backward rather than forward.
The optimistic pitch for all this new plumbing is that if one AI model gets pulled or goes bad, you just swap in another. That's true only if a comparable competitor exists to swap to. This week that got harder to take for granted: government decisions, not just engineering, can now knock a top AI out of reach on short notice.
Key takeaways
- Imprint went from ~6 deploys/week to 200-400 under 100% Claude Code/Cursor adoption in two months.
- OpenAI is acquiring Ona to give Codex persistent, customer-controlled cloud execution; Codex now has 5M weekly users.
- Databricks open-sourced Omnigent (Apache-2.0), a meta-harness above Claude Code, Codex, Cursor, Pi, and custom agents.
- US export-control directive suspended Anthropic Fable 5/Mythos 5; Artificial Analysis's frontier chart moved backward for the first time.
- Artificial Analysis swapped SWE-Bench Pro for DeepSWE in its Coding Agent Index over benchmark gaming.
- Agentic coding is producing real throughput gains — one company went from ~6 to 200-400 weekly deploys in two months.
- Tooling that abstracts across coding agents (Claude Code, Codex, Cursor) is arriving; multi-vendor design is now standard practice.
- OpenAI's Codex hit 5M weekly users and is gaining persistent cloud execution via the Ona acquisition.
- US export controls suspended Anthropic's leading coding models overnight — vendor concentration is now a policy risk, not just commercial.
- Frontier capability moved backward for the first time this week; assume swap-in parity is not guaranteed.
- Companies are handing real work — like shipping code — to AI, and shipping far more, faster.
- OpenAI's coding assistant now has roughly five million weekly users and is gaining staying power.
- A US export order abruptly cut off a leading Claude model for some customers overnight.
- For the first time, a widely-watched tracker showed top AI capability moving backward.
- The 'just switch to another AI' backup plan only works if an equally good rival exists.
Editor's Picks
The most consequential items of the week
Revised Rules of Engineering Leadership
A year's ground truth from running Imprint under near-total Claude Code/Cursor adoption: deploys went from ~6/week to 200-400/week, a mono-repo consolidation 95% driven by one engineer, full frontend static-typing done by a single engineer in weeks. Larson's argument is that domain-owning teams matter more, not less, in this regime — first-pass code is nearly free but working code still depends on your dev harness, and executive decision-making bandwidth is now the pace-limiting resource. The most useful practitioner post of the week.
The most grounded practitioner account this week comes from Imprint's engineering leader, writing after a full year of near-total AI coding adoption. Deploys jumped from roughly 6 per week to 200–400; a mono-repo consolidation and a full frontend typing overhaul were each handled by a single engineer in weeks. His takeaway isn't that you need fewer engineers — it's that domain ownership and executive decision-making have become the actual bottlenecks. Generating code is nearly free now; knowing what to build and having the systems to ship it safely still aren't.
A real company — Imprint, a fintech startup — spent a year letting AI write almost all of its code. The results are striking: software releases jumped from about 6 a week to as many as 400, and work that used to need whole teams got done by one or two people in weeks. The lesson from the person who ran the experiment isn't that developers become unnecessary — it's that experienced engineers who truly understand the product matter more than ever, because the hard part is no longer typing code, it's knowing what to build and whether it actually works.
OpenAI acquires Ona to add persistent cloud execution environments to Codex
Codex hit 5M weekly users, up 400% YTD, and OpenAI is now acquiring Ona to remove the constraint that agent sessions die with the laptop that started them. Ona's stack runs execution inside a customer's own cloud environment with scoped credentials, activity logging, and enterprise security boundaries. The acquisition reads as OpenAI conceding that Codex is being used for hours-to-days work, not minutes, and that the model isn't the bottleneck for enterprise deployment — the execution substrate is.
OpenAI's coding agent, Codex, has reached 5 million weekly users — up 400% year to date — and OpenAI is now acquiring Ona to fix the next constraint: agent sessions that die when the laptop running them closes. Ona lets agents execute inside a customer's own cloud environment with scoped permissions, full activity logging, and enterprise security controls. The acquisition signals that OpenAI's enterprise customers are running agents for hours or days, not minutes, and that the gap holding back broader adoption isn't model quality — it's the infrastructure around it.
OpenAI's coding assistant (called Codex) has grown to 5 million users a week — quadrupling so far this year — and companies are now using it for work that takes hours or days, not just quick questions. OpenAI is acquiring a company called Ona to handle a practical problem that came with that growth: when you close your laptop, any AI task in progress dies. Ona's technology lets those tasks keep running securely in the background, which is roughly what enterprise customers have been demanding.
Introducing Omnigent: A Meta-Harness to Combine, Control and Share Your Agents
Databricks' framing for Omnigent is worth reading alongside the repo. Matei Zaharia and colleagues argue the frontier of agent engineering has moved from single-model, single-harness setups (Harvey's advisor-worker pattern, Anthropic's lead-agent/subagent research system, Databricks' own Genie) to loops that span harnesses and models. The three problems they identify at the meta-harness layer — composition, contextual policies that track dynamic session state, live collaboration — are the ones a single-harness product cannot solve without becoming a walled garden.
Databricks researchers argue that AI agent deployments are maturing past the stage where a single tool handles everything. The next generation of enterprise agent work spans multiple AI systems operating together — and the hard problems are now about how those systems hand off work, enforce consistent security rules across a live session, and allow teams to collaborate in real time. A single vendor's walled-garden product can't solve this without locking you in. That's the gap Omnigent is designed to address.
Researchers at Databricks argue that using AI has evolved past asking one AI assistant a question, and that the cutting edge is now networks of AI tools handing tasks between each other. The real unsolved problems aren't about which AI is smartest — they're about how you safely manage handoffs, keep track of what each AI is allowed to do mid-task, and let multiple people work on the same AI-driven process at once. This is the thinking behind a new open tool called Omnigent.
omnigent-ai/omnigent — Meta-harness to orchestrate Claude Code, Codex, Cursor, and custom agents without rewrites
Apache-2.0 meta-harness sitting above Claude Code, Codex, Cursor, OpenCode, Hermes, and Pi, plus custom YAML agents. Sessions follow you between terminal, browser, and phone; teammates can co-drive a running session or fork it. Policies (approve-on-shell, per-session budget caps, network egress transforms that swap secrets on approved requests) attach at server, agent, or session level and check every action. Sandboxing uses bwrap on Linux and seatbelt on macOS. The design bet: harnesses will keep multiplying, but the coordination layer above them can be shared.
Omnigent is an open-source coordination layer that sits above today's major AI coding agents — Claude Code, Codex, Cursor, and others — unifying them under one interface. A session started on a terminal can be picked up in a browser or on a phone; teammates can join or fork a live session. Security controls attach at the organization, agent, or session level and check every action before it runs. The design premise: the number of agent tools will keep growing, so the governance and coordination layer above them is worth sharing rather than rebuilding inside each product.
Omnigent is a free, open-source tool that sits above multiple AI coding assistants — think of it as a universal remote for AI tools that currently don't talk to each other. You can start a task on your computer, hand it off to a colleague on their phone, or run several AI tools on the same job simultaneously. It includes built-in guardrails: spending limits per task, controls over what the AI is allowed to do, and security boundaries so that AI tools can't reach parts of a system they shouldn't.
[AINews] Fable and Mythos officially under US export controls — Latent Space
Latent Space's AINews aggregation of the Friday-night suspension: a US export-control directive forcing Anthropic to revoke Claude Fable 5 and Mythos 5 access for foreign nationals, with knock-on disruption for all users. Artificial Analysis's Intelligence Frontier chart moved backward for the first time; Cognition/Devin and Agent Arena pulled the model from their stacks. Anthropic disputes the underlying capability claim, calling it a 'misunderstanding' based on verbal, narrow evidence. The precedent — one directive taking a top-of-leaderboard model offline overnight — is the story.
Friday night, a US export-control directive forced Anthropic to revoke access to two of its top-performing models — Claude Fable 5 and Mythos 5 — for foreign nationals, disrupting users broadly and causing at least two AI platforms to pull the models from their stacks entirely. Anthropic disputes the underlying capability assessment that triggered the action. Whatever the merits, the precedent matters: a single government directive took a top-ranked model offline overnight. For any business running critical workflows on a single frontier model, vendor availability is now a policy risk, not just a technical one.
On a Friday night, a US government export-control order forced Anthropic — the company behind the Claude AI — to cut off access to two of its most capable AI models for users outside the US, essentially overnight. Several companies that had built products on those models had to scramble. Anthropic says the government's reasoning was based on a misunderstanding of what its AI can actually do. Regardless of who's right, the episode is a wake-up call: access to AI tools you depend on can be switched off by policy decisions with little warning.
Product Launches
Databricks open-sourced Omnigent under Apache 2.0, framing it as the layer above single-agent harnesses. The post from Matei Zaharia and colleagues argues that engineering practice is moving from prompting a single agent to designing loops of multiple agents across models and harnesses — and that composition, contextual security policies that track dynamic session state, and live-shared sessions are the shared problems worth solving above the harness rather than inside one.
Databricks' Omnigent launch comes with a pointed strategic argument: enterprise AI work is moving from single-agent tools to systems where multiple agents, across different platforms, operate together. The shared problems — how agents hand off work, how security policies follow a session, how teams collaborate in real time — are better solved at a shared layer than rebuilt inside every vendor's product.
Databricks, a data and AI company, released Omnigent as a free tool alongside a plain argument: the hard problems in AI right now aren't about making a single AI smarter, but about safely connecting multiple AI tools together, managing what they're each allowed to do, and letting teams collaborate on tasks those tools are handling.
- Introducing Omnigent: A Meta-Harness to Combine, Control and Share Your Agents — At Databricks, we use and build agents extensively, from coding with them at scale to shipping agent products like Genie. But even though the capabilities of agents have gotten much better, working with them feels clunky. As users, we often have 4-5 agents open at once (coding agents, Gemini search, etc) and spend our time copy-pasting text between them and Docs, Slack, and other collaboration tools. And as agent builders, we’re on a treadmill to improve our agents by combining the latest harnesses, SDKs and models. The problem is that LLM capabilities are wrapped into an agent harness, and these harnesses have different interfaces that make combining them or swapping them difficult. So we built Omnigent: a meta-harness that sits above the agents you already use (Claude Code, Codex, Pi, or custom agents) and makes them interoperable parts of a richer system. Omnigent targets the problems where a single harness stops: it adds easy ways to compose multiple agents, control them with advanced policies, and collaborate live with teammates.
Open Source
Omnigent is the highest-signal repository in the briefing: an Apache-2.0 meta-harness above Claude Code, Codex, Cursor, OpenCode, Hermes, and Pi. YAML-defined agents port across harnesses with a one-line change, sessions are addressable from terminal, web, and phone, and OS-level sandboxing (bwrap on Linux, seatbelt on macOS) ships in the default install. Policies stack server-wide, per-agent, and per-session, with stricter session rules checked first.
Omnigent is the week's most practically significant open-source release: a freely licensed coordination layer that sits above Claude Code, Codex, Cursor, and other major AI coding agents. It lets sessions move across devices, teams to collaborate on running agent workflows, and organizations to enforce security controls consistently — all without being locked into any single vendor's ecosystem.
Omnigent is a free, openly available tool designed to wrangle multiple AI assistants at once — letting them hand tasks between each other, keeping sessions alive across devices, and enforcing rules about what each AI is allowed to do. It's a sign of where things are heading: less "ask one AI a question", more "coordinate a team of AIs on a longer job".
- omnigent-ai/omnigent — Meta-harness to orchestrate Claude Code, Codex, Cursor, and custom agents without rewrites — Omnigent positions itself as an interoperability layer above individual coding-agent harnesses: swap between Claude Code, Codex, Cursor, or custom agents without rewriting orchestration logic, with policy enforcement and sandboxing built in. At 6,964 stars it's the highest-signal multi-agent framework candidate in the briefing. The value proposition sharpens as teams start mixing models (e.g. GPT-5.6 Sol for planning, Luna for execution) rather than committing to a single provider.
Industry News
Two acquisitions and a suspension. Will Larson's post from Imprint is the most useful practitioner document of the week — a year of experiments in what an engineering org looks like when close to 100% of PRs are agent-drafted. OpenAI's acquisition of Ona targets the constraint that Codex sessions end when a laptop closes. The Fable/Mythos export-control suspension, reported through Latent Space, is the reminder that upstream vendor availability is now a policy variable.
Three stories that converge on a single theme: the limits of AI at enterprise scale are increasingly organizational and infrastructural, not model-quality. One practitioner shows what happens to an engineering org a year into near-full AI adoption. OpenAI's acquisition of Ona tackles the problem of agents that can't outlast a closed laptop. And an overnight export-control suspension of Anthropic's top models is a sharp reminder that upstream vendor availability is now a policy variable your procurement team needs to account for.
This week's industry news covers what AI is actually doing inside real companies right now — including a striking experiment in near-total AI coding, a telling acquisition by OpenAI, and a government order that abruptly cut off access to some of the most powerful AI tools available, reminding everyone that these services can disappear overnight for reasons that have nothing to do with technology.
- Revised Rules of Engineering Leadership — Will Larson documents five revised rules for engineering leadership in an AI-tool-augmented hypergrowth context, grounded in a year of concrete experiments at Imprint. Key claims: (1) complex migrations can now be driven by a single engineer in ~10% of the former time; (2) 1st-pass code is cheap but working code still depends on your development harness; (3) the base case of most processes should now be optimized for autonomous agent execution; (4) durable, high-ownership teams matter more than ever because domain context is the scarce resource; (5) fast, binding decision-making is a prerequisite to actually capturing the pace benefit. Includes detailed project examples: 200–400 deploys/week (up from 6), 100% Claude Code/Cursor adoption in two months without mandate, mono-repo migration done by one engineer.
- [AINews] Fable and Mythos officially under US export controls — Latent Space — Latent Space AINews covering the US government's export control directive on Anthropic's Claude Fable and Claude Mythos models, forcing suspension of access for foreign nationals and international users. The event is referenced in the Gray Swan / Latent Space red-teaming episode as the proximate cause for AI security suddenly becoming the "talk of the town," and is cited by Nathan Lambert at Interconnects as the "starting gun of a new era in AI governance." Documented separately because it functions as a cross-referenced event across multiple high-signal posts this week.
- OpenAI acquires Ona to add persistent cloud execution environments to Codex — OpenAI announced the acquisition of Ona, a cloud execution infrastructure startup whose technology provides secure, persistent, customer-controlled cloud environments for long-running agents. Codex now has 5M+ weekly users (up 400% YTD) but is constrained by session-bound execution; Ona's stack removes that constraint, allowing agents to continue work when the originating laptop is closed. The Ona team will join OpenAI's Codex team post-close, with the specific goal of enabling production-grade enterprise agent deployments with scoped credentials, activity logging, and security controls inside customers' own cloud perimeters.