News · · 9 min
Period: September 12 — morning of September 13.
There are almost no new frontier models today. A more interesting shift is happening at a higher level: AI teams are beginning to treat harness, evals, permissions, and infrastructure efficiency as distinct engineering disciplines.
This is significant material from the previous window, worth adding as a backfill.
What happened. OpenAI revealed the architecture of Habitat — the online-storage layer serving ChatGPT, Codex, and the API. The old Python service at peak handled over 20 million requests per second. In Q2 2026, two engineers, using Codex and GPT‑5.5, rewrote it in Rust; the new version already serves about 95% of production traffic.
What the data showed.
According to OpenAI's measurements, the Rust version is:
The entire Habitat already operates at a scale of over 500 PB of data and more than 70 million storage requests/sec at the underlying layer.
This is OpenAI's production measurement, not a benchmark.
Why this is important. At the scale of AI products, the efficiency of ordinary software is once again becoming strategic.
The causal chain:
more users and agents
→ more state/tool/storage operations
→ more expensive infrastructure
→ ordinary systems engineering begins to significantly impact AI unit economics.
Not every gain requires a new model or GPU.
What you need to understand now: tail latency, CPU efficiency, memory footprint, online storage, Rust.
What this could change. If an AI product is already large, you should optimize not only token cost. Storage, serialization, tool infrastructure, queues, and orchestration can contain orders of magnitude larger efficiency reserves.
Publication date: September 11. Primary source: OpenAI — Scaling online storage to over 1 billion ChatGPT users
What happened. On September 12, Anthropic CEO Dario Amodei publicly called for slowing the growth of frontier model capabilities so that safety engineering can keep up. The practically more important part of the statement — Anthropic is ready to provide permanent independent evaluators with access equivalent to internal risk teams, including the ability to publish findings without Anthropic's editorial control.
The proposed scheme consists of three levels:
independent evaluation inside labs
→ shared safety standards between companies
→ interstate coordination.
Why this is important for engineering. Until now, external model evaluation usually occurred after release or on specially prepared checkpoints.
If embedded third-party evaluation becomes the norm, eval infrastructure must become part of the development lifecycle itself:
training → capability eval → agent eval → red team → deployment gate → continuous monitoring.
That is, evals are becoming closer in role to CI/security testing, not a marketing benchmark table.
What the data showed. This is a policy/engineering proposal, not an experiment. Amodei's stated predictions about the capabilities of future agent swarms are assessments from the head of Anthropic, not a confirmed benchmark.
What you need to understand now: capability evaluation, red teaming, deployment gate, continuous evaluation, frontier model.
What this could change. For an ordinary product team, there are no direct actions yet. But the principle is useful already: the person who writes an agent workflow should not be the only person who determines if that workflow is safe.
Date: September 12. Additional sources: Reuters and AP.
What happened. Claude Code 2.1.269 introduced claude plugin eval: a plugin developer can now run a reproducible evaluation suite and get scored results in JSON and HTML.
The release also added repository metadata to OpenTelemetry and the ability to see diffs of files changed by a Bash-tool.
Why this is important. Skills and plugins are quickly becoming a separate software layer on top of models, but until now they were often developed almost like prompt engineering:
wrote instructions → tried manually → seems to work.
Now a normal lifecycle is emerging:
skill → eval cases → scores → regression testing → release.
This is especially important because the same model with different harnesses already shows significantly different results on agent benchmarks.
What the data showed. Anthropic does not publish a general benchmark of quality improvement. The news is that a reproducible evaluation mechanism for the plugin layer itself has appeared.
What you need to understand now: Agent Skills, eval suite, regression testing, OpenTelemetry, plugin harness.
What this could change. If a company creates its own skills/instructions for coding agents, they should be stored as a versioned software artifact and regression-tested before updating, rather than manually changing the production prompt.
Release date: September 11; the significant developer signal continued to spread on September 12. Primary source: Anthropic Claude Code releases
What happened. Reuters and CuttingRoom integrated the Reuters MCP server into the browser-based video editor ShortCut. The editor can use natural language to search Reuters footage, add material to the timeline, and perform editing operations — cutting, captions, audio mixing, colour correction, and formatting.
At the same time, the workflow preserves newsroom-specific editorial rules, and Reuters footage can be mixed with the editorial team's own content.
Why this is important. MCP is gradually moving out of developer tooling into vertical business workflows.
The architecture is becoming:
AI interface
→ MCP
→ trusted proprietary content/data
→ domain application
→ human-controlled final artifact.
In this case, the value of AI lies not so much in the model, but in access to verified Reuters content and the existing editing workflow.
What the data showed. Reuters has not yet published quantitative data on time savings or newsroom productivity. This is a product integration, not a performance study.
What you need to understand now: MCP, tool integration, provenance, human-in-the-loop, vertical AI.
What this could change. For SaaS companies, a good question is whether they should turn their own API into an agent-accessible tool/MCP layer, so that the product is used not only by people through a UI, but also by agents within other workflows.
Date: September 12. Primary source: Reuters — CuttingRoom integration
Among rapidly gaining attention projects, milind-soni/OpenMausBot stands out: an open-source shell where Claude, Codex, or Grok are presented as separate "contacts," each with its own context, model, apps, and computer environment. At the time of checking, the repository has about 2.5k stars and is actively updated.
The architecture is more interesting than the star count.
Each bot can get:
Permissions for critical actions remain a separate layer, and credentials should not be passed to the UI.
Signal. The architecture, which just a few months ago was only shown by commercial agent products, is quickly becoming reproducible open source:
model → harness → isolated computer → apps → approval broker → persistent tasks.
What you need to understand now: computer-use agent, permission broker, VM isolation, MCP, persistent agent.
What this could change. For now, this is an early project, not a production recommendation. But teams should watch this category: personal/corporate agent runtime is becoming a separate open-source software class.
stablyai/orca currently has about 32.8k stars. But the feature set is more important: parallel Git worktrees, several different coding agents simultaneously, remote agents via SSH, mobile monitoring/steering, and human review of AI-diffs.
This is a continuation of the signal from previous issues, so the project itself is not new news. What's new here is the consolidation of the category:
IDE for a human
is gradually turning into
control plane for multiple working agents.
What you need to understand now: parallel agents, Git worktree, agent fleet, human supervision.
What this could change. When internally testing multiple coding agents, it makes sense to measure not only the best result of each, but also the overhead of the human who has to manage parallel attempts.
On September 12, OpenAI and AI Tinkerers held Agents, Everywhere — a synchronous agent-focused build day in about 50 cities. The ecosystem involves OpenRouter, CopilotKit, Exa, Trigger.dev, Auth0, Mozilla, and other tools at various levels of the agent stack.
The hackathon itself is not a technological breakthrough. The composition of the infrastructure is interesting:
model/API → routing → search → auth → orchestration → channel/UI.
That is, the developer ecosystem already perceives an agent not as a single library, but as a full-fledged composable stack.
Today, the most important thing is not a new model, but the professionalization of the layer around the model.
OpenAI shows that systems engineering can yield 6× CPU and 15× memory efficiency. Anthropic is making evals a separate lifecycle layer. Claude Code allows regression-testing plugins. Reuters uses MCP as a production integration protocol. Open-source projects are building permission brokers and separate computers for each agent.
The market is gradually shifting from the question "which model to choose?" to the question "how well is the system around the model designed?"
Do we regression-test prompts, skills, and plugins the same way as regular code, or are harness changes still rolled out manually?
What portion of our AI product's cost lies outside model inference — storage, orchestration, serialization, queues, browser/VM, and network?
Could our product become a tool/MCP endpoint for others' agents, not just an application with a human UI?
If we run several coding agents in parallel, do we have a control plane for state, permissions, results, and human review?
How independent is the person or team evaluating the safety of an agent workflow from the people who developed it?