AI News Roundup: November Five and the Agent Land Grab

Overview: Thursday was the day the agent layer stopped being experimental and started getting institutionalised. Google handed its agent runtime to the Cloud Native Computing Foundation, Meta bought the tooling startup that was quietly becoming the standard place agents do their work, and Perplexity put always-on agents behind a $200 monthly subscription. Underneath the product noise, the balance sheets moved too: OpenAI's president said the company could be IPO ready in 2027, and Agility Robotics began trading with humanoid labour priced as a service contract.

Google Donates Its Agent Runtime to the CNCF and Ships 1.0

Google put its agent runtime into a foundation and cut a 1.0 release on the same day, which is the closest thing this market has to a standards moment. The practical effect is that the scheduler, session handling and tool-permission layer stop being a vendor differentiator and become shared infrastructure, the way Kubernetes did for containers. For teams that have spent a year writing glue code against a single vendor's agent API, the calculation changes: the portable surface just got larger and the lock-in argument got thinner. The 1.0 tag also matters for procurement: "donated to a foundation, version 1.0" survives a security review.

Meta Buys Conductor, and the Cloud Shuts Down in Two Weeks

Meta acquired Conductor, the agent workspace tool, and gave hosted users a two week shutdown window. It is the clearest sign yet that agent runtimes are being consolidated rather than competed on: the buyer was not looking for another model, it was looking for the surface where people already keep their agent sessions and project state. Two weeks is short enough to be a forcing function, and anyone who built on Conductor is now doing a migration evaluation they did not schedule. The lesson repeats itself every cycle: if your agent's memory lives in someone else's startup, that memory has a migration date.

Perplexity Launches Its Own Take on AI Agents for Work

Perplexity shipped Teammates, always-on agents that pull project context, skills and memory from the tools a workspace already uses, and each one gets its own thread plus a Slack presence. The pitch is that an organisation ends up with hundreds of agents working for each person, which is either a productivity story or a supervision story depending on your org chart. Pricing is the sharper signal: Teammates sits behind the $200 per month Max plan, so this is positioned as headcount-adjacent software, not as a cheap add-on. That puts Perplexity in the same bracket as OpenAI's and Anthropic's agent products, and makes auditing an always-on agent the buyer's problem.

OpenAI President Says the Company Could Be IPO Ready in 2027

Greg Brockman told a Wall Street Journal conference that OpenAI would be ready to go public in 2027, and the numbers behind the answer are more interesting than the date. He described the first half of 2026 as rough, admitted the company was not even in the agentic coding conversation, and then pointed to roughly 70 percent quarterly growth in annualised revenue in the third quarter driven by enterprise doubling. Enterprise is now the majority of revenue, up from about 30 percent, with JPMorgan Chase as the marquee logo. Read that against Anthropic's filing numbers and the shape of the next two years is clear: both labs are converting model advantage into enterprise contracts and multi-decade compute commitments, and both will need public markets to carry the difference.

Agility Robotics Goes Public and Puts a Price on Humanoid Labor

Agility Robotics listed, which makes it the first credible public comp for humanoid labour. The interesting number is not the valuation but the unit economics it drags into daylight: robots are now a line item competing with staffing budgets, and that comparison is what a public market will price quarter by quarter. Humanoids that were pitched as platform science now have to justify themselves as cost-per-shift, with maintenance, uptime and integration labour included. A robot that rents itself out by the hour is an agent with a body and a billing meter.

What to Watch Tomorrow

• Meta warns the midterm backlash will slow data centers: Meta president Dina Powell McCormick said Democratic wins make construction harder and conceded the industry failed to explain the benefits, with five of its 23 state sites planned at a gigawatt or more. (TechCrunch)
• Gemini Omni 1.2 Flash cuts video pricing 30 percent: Google made 15-second native clips generally available at $0.07 per second for 720p, with multi-turn conversational editing and SynthID on every output. (Google)
• The political bill for the buildout: coverage of how the blue wave lands on AI policy is only starting. (The Verge)
• Cyber capability keeps outrunning its guardrails: the day's GPT-6 Cyber analysis shows offensive security models now work, and the containment story is written after deployment.

Builder Impact

• Standardise at the runtime, differentiate at the workflow. With a CNCF-hosted agent runtime at 1.0, portability is no longer a nice-to-have argument you make to a client. Build your orchestration so the runtime is swappable and your value sits in the tools, data and evaluations you own.
• Treat agent memory as an asset with a migration risk. Conductor's two week shutdown is the template. Assume every hosted agent workspace can disappear and keep exportable state from day one.
• Price supervision, not just agents. Perplexity's $200 tier and hundreds-of-agents framing both assume someone watches the output. Budget for review, logging and rollback as first-class components, or the labour savings get eaten by the audit.
• Enterprise is the revenue, not the demo. OpenAI's enterprise majority and Anthropic's compute commitments mean procurement cycles, security reviews and SLAs now shape model roadmaps. Read release notes for compliance language, not just benchmarks.
• Model the physical cost. Power, permitting and public-market scrutiny on robot labour decide which agent products can actually scale. Build assumptions that survive a gigawatt queue and an earnings call.