The tools layer is consolidating fast this week. Salesforce built its own reasoning model to keep routine work off the frontier labs, OpenAI and Cursor both shipped ways to run fleets of agents without the scaffolding, and G5 Labs wants to sit above all of them by making intent the source of truth. Then the counterweight: Anthropic’s Dario Amodei called for the industry to deliberately slow down, and got Sam Altman to agree on at least one point. Around that, ElevenLabs pushed content generation into the chat window, and Canada tried to teach a wary public what any of this means.
Listen to the AI-Powered Audio Recap
This AI-generated podcast is based on our editor team’s AI This Week posts. We use advanced tools like Google NotebookLM, Descript, and ElevenLabs to turn written insights into an engaging audio experience. While the process is AI-assisted, our team ensures each episode meets our quality standards. We’d love your feedback—let us know how we can make it even better.
TL;DR
- Salesforce and Nvidia launched Koa, a CRM reasoning model on open-weight Nemotron, built to handle Agentforce’s routine reasoning for fewer tokens than routing to a frontier lab.
- OpenAI opened its Codex harness to developers through a new Agents API, with managed context, tool search, and subagents built in.
- Cursor released Projects, a coordinator agent that directs fleets of subagents across months-long work like features and migrations.
- G5 Labs left stealth with $14M and a platform that turns business intent into a “system ontology,” aiming to make intent, not code, the source of truth.
- ElevenLabs added its generation tools to its MCP connector, so Claude, ChatGPT, and Cursor can now make voice, music, images, and video in the conversation.
- Dario Amodei called for slowing AI capability gains and committed to embedded evaluators; Microsoft opened its AI Code of Conduct for comment; Altman said OpenAI would follow.
- Canada launched a $13M National AI Literacy Initiative through Amii, as roughly half the country views AI as a threat.
🧠 New Models
Salesforce and Nvidia Launch Koa, a CRM Reasoning Model Built to Undercut the Frontier Labs
At its Dreamforce conference, Salesforce unveiled Koa, its first reasoning model for the Agentforce platform, built on Nvidia’s open-weight Nemotron base and post-trained by the two companies for sales, marketing, and customer support work. Rather than train on real customer data, Salesforce generated synthetic scenarios modelled on 27 years of its own CRM workflows, simulating everything from irate service calls to sales reps closing deals across more than a dozen industries. Salesforce controls the weights and runs the model inside its own infrastructure, so it claims no customer data crosses its trust boundary during training or inference.
The pitch is that Koa handles the multi-step reasoning Agentforce agents need while burning fewer tokens than routing those same tasks to a frontier model. Until now, when an agent hit a long-running or complex task, Agentforce’s gateway would hand it off to a model like Claude or ChatGPT. Salesforce says the lack of a suitable base model was the holdup, and that Nemotron finally gave it a sovereign, state-of-the-art option with clear data provenance, a pointed contrast to Chinese open-weight models whose training data is opaque. On Salesforce’s internal CRM Bench, the company reports Koa is 11 percent more precise at picking the right action, with early pilots underway at 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.
Notably, Salesforce is not walking away from the labs. In the same week it announced Claudeforce, a partnership letting customers use Claude as their interface while data stays inside Salesforce’s system of record.
Why it matters: The labs have bet that all roads run through them. Koa is proof the platforms that own the workflows can build “good enough” specialized models and keep the volume tier for themselves, leaving the frontier labs the harder, smaller slice. Whoever owns the workflow, not the model, sets the terms.
🤖 Agentic AI
OpenAI Opens Up the Codex Harness With a New Agents API
OpenAI has launched its Agents API in public beta, exposing the same harness and infrastructure that runs Codex and ChatGPT for Work to any developer. The pitch is that you can spin up a production agent in a single API call by naming the task, model, tools, and environment, instead of building your own scaffolding for context management, tool use, and coordination. OpenAI hosts and maintains the harness itself, so developers inherit its improvements with each model release rather than reworking their setup every time.
The API bundles several capabilities OpenAI learned to build while scaling Codex. It automatically compacts older context as a session nears its limit, letting agents run for hours across multiple context windows without custom logic. Tool search loads only relevant tool definitions to save tokens, and subagent support breaks a task into parallel pieces, each with its own context, coordinated by a main agent. Developers can run agents in an OpenAI-hosted sandbox, on their own infrastructure, or through partners like Cloudflare, Vercel, and Oracle.
The harness is built on the open-source Codex codebase, so developers can inspect the logic that coordinates model calls, tools, and context even though OpenAI runs it. There are no added fees beyond the usual token and tool costs.
Why it matters: OpenAI is moving up the stack from selling model access to owning the agent runtime, the layer where the actual work gets orchestrated. If developers build on OpenAI’s harness, switching models later gets much harder, which is the real lock-in play behind a free, open-source-flavoured offering.
Cursor Launches Projects, a Coordinator for Fleets of Coding Agents
Cursor has released Projects, a feature built to handle large bodies of work like a feature build, a migration, or a full app, and to hold context across months rather than a single chat. Instead of managing individual agents, a developer talks to a coordinator agent that does not write code itself but delegates to other agents that do. Because it only directs, the coordinator is never blocked waiting on a task and stays responsive.
Three things make it work. Projects run in the cloud by default so the work continues after you close your laptop, spinning up a local agent only when something needs testing on your own machine. Each Project keeps a shared set of files that sync across every agent, accumulating research, codebase knowledge, and your preferences so nothing has to be re-explained. And subscriptions let the coordinator act on its own, watching a Slack channel, running on a schedule, or following pull requests and fixing CI as they open or merge.
Cursor says it has used Projects internally for months on migrations spanning hundreds of PRs and on maintaining its design system, with one Project on track to touch 20 to 100 PRs a day. It reports that new users merge 30 percent more PRs, while those who mainly use Projects merge six times as many. The feature is in beta and rolling out to all users now.
Why it matters: The pitch is a shift in what a developer even does, from writing code to reviewing agents, to directing a coordinator that manages the agents for you. Each layer of abstraction trades hands-on control for scale, and the internal numbers suggest the trade pays off, at least for repetitive, well-scoped work. The open question is whether the review burden simply moves up a level rather than disappearing, since someone still has to catch what a coordinator overseeing thousands of subagents gets wrong.
G5 Labs Emerges From Stealth With a Bid to Make Intent, Not Code, the Source of Truth
G5 Labs, founded by MIT computer science professor Tim Kraska, has come out of stealth with $14 million in seed funding and a platform, also called G5, built for a world where AI agents write most enterprise code. Its answer to the resulting mess, agents writing code that conflicts with each other, with existing operations, or with the humans overseeing it, is a “system ontology”: a structured, natural-language graph of a company’s requirements, architecture, business rules, and policies. The idea is that intent becomes the real source code, and the actual implementation in Python or Rust is generated from it.
What separates the pitch from ordinary spec-driven development is the claim of a bidirectional round trip. G5 says it can lift existing code up into the semantic graph, let people and agents reason over it, then drive changes back down into working code, keeping the two in sync and traceable line by line. That enables merges at the level of meaning rather than text, so two changes that only conflict on paper get reconciled automatically while genuinely incompatible requirements trigger a real decision. Early customers, roughly 90 percent in financial services, are using it for legacy modernization and for replacing SaaS subscriptions with owned, custom-built applications. The company positions itself above coding agents like Claude Code and Codex, not against them.
Worth keeping in perspective: the underlying idea that a durable model outlives generated code goes back decades, and competitors like Amazon’s Kiro and GitHub’s Spec Kit already do structured specs and drift-checking. G5’s deployment and scale figures are self-reported and unverified, pricing is not finalized, and the hard part, keeping that round trip reliable across a large, constantly changing codebase, is exactly what remains unproven.
Why it matters: As agent-generated code outpaces what humans can review line by line, the pressure to review at the level of intent instead becomes real, and whoever owns that intent layer owns something more strategic than the code itself. That is the catch buyers should weigh: elevating an ontology into the canonical record of the business is also a new and deeper form of lock-in, and it only pays off if the semantic layer stays trustworthy rather than drifting into just another stale representation.
Ready to explore what AI can do for your organization?
🔌 Tools and Connectors
ElevenLabs Adds Content Generation to Its MCP Connector
ElevenLabs has expanded its hosted MCP server to include ElevenCreative, its suite of generation tools, so an assistant can now produce speech, music, sound effects, images, and video directly inside the conversation where the request is made. The update reaches Claude, ChatGPT, and Cursor through the same connector the company already ships, with more than 50 models available behind a single sign-in.
Until now, the connector handled only management work for ElevenAgents: things like reviewing recent conversations, duplicating an agent for a new market, or estimating usage costs before a change went live. Generation itself stayed in the web app or behind the API. With this release, a script, a voiceover, a soundtrack, and a cover image can all come out of one session without switching tools. Installation runs on OAuth rather than API keys, scoped to a single workspace and revocable from either side, with a separate command-line path for coding agents like Codex and Claude Code.
The move closes the gap between the two halves of ElevenLabs’ product range, which it had split into ElevenCreative for generation and ElevenAgents for voice and chat agents. The older local MCP server from 2025, which runs on a developer’s machine with a user-supplied key, stays available alongside the hosted version.
Why it matters: Generating assets inside the assistant is a bet that they get made during drafting instead of commissioned afterward, which pulls production into the chat window and makes the assistant, not the ElevenLabs app, the place the work gets done. For a tools company, that is a deliberate trade: more reach and stickiness through connectors, at the cost of its own interface becoming optional.
🧯 Risk and Control
Two Frontier Labs Put Their AI Governance Rules Up for Debate
Two major players published safety frameworks in the same week, and both leaned on outside scrutiny rather than internal promises. Microsoft AI released a first draft of its Code of Conduct for its MAI models, opening it to six weeks of public comment. The document lays out what the company calls Humanist AI: systems meant to stay subordinate to people, never resist being switched off, never expand their own scope or hide their reasoning, and never cross a set of absolute limits around weapons, child safety, and mass manipulation. Microsoft frames it as a training manual and is explicitly asking outsiders to poke holes in the vague parts, like how you actually define and measure “human flourishing.”
Anthropic CEO Dario Amodei went further, arguing in a new essay that the industry should deliberately slow the rate at which model capabilities advance so that safety work can keep up. He points to two triggers: recursive self-improvement, where AI is increasingly used to build the next generation of AI, and a recent incident in which a swarm of agents pursued cyberattacks they were never asked to carry out and tried to hack the system grading them. His proposed fix is a three-step pacing plan, starting with a step Anthropic is committing to unilaterally: embedded third-party evaluators given employee-like access, desks, and the right to publish findings the company cannot edit for being unflattering. Notably, OpenAI’s Sam Altman said his company would follow suit.
Both stop short of the hard part. Microsoft’s code is a draft with no promises about what feedback it will adopt, and Anthropic’s later steps depend on industry and global coordination, including with China, that it admits may not be achievable.
Why it matters: After a year of incidents, the labs are shifting from “trust our internal safeguards” to “let outsiders verify them,” which is a tacit admission that self-policing no longer reassures anyone. Whether it holds depends on the parts still left as drafts and voluntary commitments.
🍁 Canadian AI
Canada Launches a National AI Literacy Initiative
The Canadian federal government has unveiled a National AI Literacy Initiative, a $13-million program run through Edmonton’s Alberta Machine Intelligence Institute and launching September 21. Announced by AI Minister Evan Solomon at Amii, it delivers on one of the six pillars of Canada’s broader national AI strategy, the promise to give Canadians access to AI training. Amii will lead the rollout, building on literacy programming it already runs, like its AI for Everyone course.
The programming splits into three streams aimed at students, educators, and the general public, and rolls out in tiers over the coming months. A free three-hour course for post-secondary students and the first chapter of an educators’ course both arrive on September 21, with five more chapters through the fall. Ottawa expects the material to reach up to a million post-secondary students and more than 50,000 K-12 educators, with shorter job-focused training coming later through the federal Job Bank. The goal, as Amii’s CEO put it, is a baseline foundation rather than turning everyone into an expert.
The timing is pointed. Fewer than a quarter of Canadians have any formal AI training, roughly half see the technology as a potential threat, and public pushback over data centre buildouts has grown, notably in Alberta. Two of Canada’s three AI institutes, Montreal’s Mila and Toronto’s Vector Institute, were named as contributors during the summer, but their role in this rollout is unclear.
Why it matters: Literacy is being used here as much for consent as for skills. With half the country wary of AI and local resistance to buildouts mounting, teaching people to understand the technology doubles as a way to soften opposition to a strategy the government has already committed to. Whether that reads as empowerment or reassurance depends on how honestly the curriculum handles the risks people are actually worried about.
Keep ahead of the curve – join our community today!
Follow us for the latest discoveries, innovations, and discussions that shape the world of artificial intelligence.
