In partnership with
> SPOTLIGHT
WHAT MATTERS TODAY
OpenAI is combining its Codex and ChatGPT strategy, with Codex seen as a stronger agent harness for tool use and action-taking. The product boundary between a coding agent and a general assistant has not disappeared, but the direction is clearer: OpenAI wants to use Codex's action layer to push ChatGPT beyond chat.
That makes Codex more strategically important than a developer tool alone. If Codex is only a coding assistant, it is a vertical product for developers. If its harness moves into ChatGPT, it becomes part of a broader system that can write, edit, test, call tools, move through multi-step work, and finish tasks. OpenAI may be treating coding agents as the best proving ground for ChatGPT's future as a work-execution system.
Google agreed to pay SpaceX $920M per month from October 2026 through June 2029 for access to roughly 110,000 NVIDIA GPUs, plus related CPUs, memory, and compute components. The deal gives SpaceX/xAI a way to monetize data centers built for Grok-related workloads as SpaceX prepares its IPO story and carries heavy AI capex.
This is less a normal cloud contract than a compute-market signal. Google has its own AI infrastructure, but demand is tight enough that one of the world's largest AI platforms is buying bridge capacity from outside. For SpaceX/xAI, expensive AI infrastructure can become leasable capacity even if Grok has not yet matched the market impact of the leading models. AI infrastructure is starting to look like a strategic asset that can be allocated, rented, and turned into a growth narrative.
Microsoft AI chief Mustafa Suleyman said the company was "set free" from OpenAI to pursue superintelligence. It is a strong positioning signal, but it does not mean Microsoft has walked away from its OpenAI partnership.
The point is Microsoft wants customers and investors to see Microsoft AI as its own roadmap, not just the enterprise distribution channel for OpenAI. The OpenAI-Microsoft relationship still matters, but the strategic overlap is rising: both companies want to own models, product surfaces, and enterprise trust. As AI becomes an execution layer inside software, the question of who really owns Microsoft's AI future moves directly into product strategy.
> SIGNAL HEADLINES
CAPTURE THE SHIFT
Claude Cowork raises usage limits: Claude says it doubled Claude Cowork usage limits for the next month so users can delegate larger and more complex tasks. Usage limits are becoming both a growth lever and a competitive signal in agent work.
Cursor ships Design Mode: Cursor introduced Design Mode, letting users point, draw, or talk to update UI. Coding tools are moving closer to multimodal product-building interfaces where visual feedback and implementation sit inside the same loop.
Codex adds search inside settings: OpenAI Developers says Codex settings can now be searched and grouped by category. Small details like settings search show Codex becoming a more configurable work environment, rather than a simple coding prompt box.
ChatGPT can send emails from writing blocks: ChatGPT on the web can now send emails directly from writing blocks without leaving the conversation. It is a small but clear move toward action-taking: the assistant does not stop at the draft, it starts touching the real communication workflow.
Companies are starting to feel the token bill: TechCrunch reports that companies are adding limits, monitoring, and ROI controls as agentic tools drive token consumption higher. AI adoption is entering its governance phase: access is the first step, but cost attribution, usage controls, and ROI discipline decide whether scaling is sustainable.
Supabase hits a $10B valuation on vibe-coding demand: Supabase raised $500M at a $10B pre-money valuation, with database launches up more than 600% year over year and more than 60% of launches coming from some AI tool. Vibe coding is repricing the infrastructure underneath AI-generated apps, far beyond the surface image of quick app demos.
Brian Chesky is planning a new AI lab: TechCrunch reports that Airbnb CEO Brian Chesky is planning to back a new AI lab. Frontier AI is expanding beyond pure research labs; founders with strong product and design instincts want a more direct role in how AI is built and turned into experience.
The new devtool moat is becoming the tool agents reach for: Nicolas Dessaigne argues that the moat in the agent era is becoming the tool agents reach for, using Supabase as an example. Devtools now need clean docs, easy-to-wire APIs, and agent discoverability alongside human developer preference.
YC launches Paxel for coding-session analysis: Y Combinator introduced Paxel, a local Docker tool that analyzes Claude, Codex, and Cursor sessions to profile how someone builds with AI. A new meta-layer is forming around agent work: observing sessions, finding bottlenecks, and improving how humans use agents.
Engineers still need to supervise coding agents: Aaron Levie argues that coding is highly automatable, but engineers are still needed to oversee agents. The work is shifting toward supervision, review, and system judgment rather than a simple story where agents ship everything alone.
AI outcomes still depend on builder choices: a16z shared Mira Murati's view that dystopian and utopian forecasts about frontier AI are too simplified because builders still have agency in how AI is built and deployed. It is a useful counterweight to deterministic AI narratives: outcomes depend on product design, incentives, and governance, not the capability curve alone.
> ONE PRACTICAL USE OF AI TODAY
Review artifacts, not every line of code
When teams use coding agents, PMs and operators can fall into two bad modes: reviewing code like an engineer, or trusting the agent too early because the prototype runs.
The better move is to review the artifact layer above the code: plans, decision logs, docs, constraints, tests, evals, source-of-truth files, and guardrails. In 30 minutes, you can set up a review loop that keeps you accountable for what ships without making you the bottleneck.
Start with a simple artifact review loop:
Start from a visible prototype: ask the agent to create something runnable before a long alignment discussion. A prototype makes requirements, edge cases, and tradeoffs visible faster.
Keep source-of-truth where the agent can read it: put strategy, requirements, constraints, API rules, and definition of done in a clear file or folder. The less the agent has to guess, the less the output drifts.
Review artifacts instead of raw code: after each task, check what changed in the plan, what decisions were added, which tests were created, which docs were updated, and which behavior was affected.
Cross-examine confident answers when the cost of being wrong is high: ask the agent for evidence, test results, unresolved risks, and assumptions. Do not debate every detail; focus on the places where mistakes are expensive.
Turn repeated failures into guardrails: if the agent keeps missing the same pattern, do not just fix it manually. Turn it into a test, eval, policy, checklist, or hook.
Enforce non-negotiables mechanically: things like secret handling, destructive commands, permission boundaries, and formatting rules should be enforced through tool permissions, hooks, or validators.
→ If behavior changed but docs or source-of-truth did not, the implementation may have drifted away from current product intent.
→ If a new decision exists without a test or eval, you are asking human memory to carry a rule that a machine should check.
→ If risks and assumptions are empty on a complex task, treat that as a reason to cross-examine, not as a sign that everything is fine.
→ If non-negotiables only live in the prompt, turn them into enforceable gates before the next task runs.
This workflow keeps PMs in control at the product and system layer instead of pulling them down into raw diffs. As agents do more work, the operator skill that matters is knowing which artifacts must be correct so the code underneath does not drift.
> PRESENTED BY MINTLIFY
AI Agents Are Reading Your Docs. Are You Ready?
Last month, 48% of visitors to documentation sites across Mintlify were AI agents, not humans.
Claude Code, Cursor, and other coding agents are becoming the actual customers reading your docs. And they read everything.
This changes what good documentation means. Humans skim and forgive gaps. Agents methodically check every endpoint, read every guide, and compare you against alternatives with zero fatigue.
Your docs aren't just helping users anymore. They're your product's first interview with the machines deciding whether to recommend you.
That means: clear schema markup so agents can parse your content, real benchmarks instead of marketing fluff, open endpoints agents can actually test, and honest comparisons that emphasize strengths without hype.
Mintlify powers documentation for over 20,000 companies, reaching 100M+ people every year. We just raised a $45M Series B led by @a16z and @SalesforceVC to build the knowledge layer for the agent era.
> WORTH READING
ANALYSIS & THESIS
TMLS argues that AI cost control has to move into the request path, where decisions about model choice, routing, context, and tool calls happen while the workload is running. The useful shift is from after-the-fact reporting to an operating model that can intervene before token spend expands.
FinOps Foundation frames tokens as the smallest useful unit for understanding AI workload value and cost. The key lens: falling token prices do not automatically mean lower total bills, because agents can use more steps, more context, and more retries until total usage still rises.
An arXiv paper finds that some agentic coding tasks can consume far more tokens than traditional code chat. It gives research backing to the cost problem engineering teams are starting to face: agent work is harder to forecast because it can expand the number of steps needed to complete a task.
