𝗧𝗵𝗲 𝗴𝗼𝗼𝗱 𝗻𝗲𝘄𝘀: anyone can create an agent now. 𝗧𝗵𝗲 𝗯𝗮𝗱 𝗻𝗲𝘄𝘀: now you have govern them all. Introducing ChiriBrain, our AI orchestration platform. Every agent, every permission, every token spend, in one place. ✔️ Set what each one can access. ✔️ See what each one costs. ✔️ Control built in from the start. All your agents. Finally under control.

Tokens are becoming the substrate of labor, and most finance teams have no real way to project what that costs a year from now. That's the line our CTO Matthew Wimberly landed on partway through a panel with @IgniteGTM on why AI bills keep climbing even as token prices fall. His point was blunt: If a company can't project its token spend years out, that's a business continuity risk. Part of why the number stays fuzzy isn't an accident. Model providers have an incentive to keep consumption opaque. It's crude but true: the first baggy is always free. Chiri builds cost controls into a client engagement from day one, so the return on a workflow gets defined before a single token is spent. If your team has ever stared at an AI bill and couldn't explain the jump, we would like to hear what that number looks like on your end.

Two new executive disciplines are emerging, and neither one had an owner on the org chart a year ago. The first is token optimization: matching the right task to the right model instead of routing everything through the most expensive one. The second is alpha protection: making sure the work that makes your company distinct doesn't quietly become someone else's training data. The market is already voting. @vercel's June 2026 AI Gateway data shows open-weight models running 29 percent of all gateway tokens, up from 11 percent in April, on under 4 percent of total spend. CNBC's July investigation found Chinese-origin models taking 30 to 46 percent of enterprise token volume on OpenRouter. Alpha protection is the harder one to see coming. When Anthropic launched a legal AI plugin in February 2026, Thomson Reuters fell 16 percent in a day, RELX fell 14 percent (its steepest drop since 1988), and Wolters Kluwer fell 13 percent. That's how the market responds when a model provider learns enough about a vertical, from aggregate usage across its customers, to compete in it directly. Neither discipline requires a chief AI officer or a six-month program. It requires your CFO, your CISO, and whoever owns your AI spend in the same room, asking which workflows can move to a cheaper model and which ones need a real contract, not just a promise, behind them. What does your organization actually know about where its token spend goes, and what it's protecting on the way there?

Ethan Mollick named the harness layer in June. Ben Thompson wrote about the same architecture in March. Neither piece was about Chiri. Both landed on the same conclusion: the model is the commodity. The layer connecting it to tools, workflows, and enterprise systems is where durable value lives. Thompson's framing at @stratechery was the Apple analogy — integrated platforms capture margin the same way hardware-software integration does. Not because the hardware is special, but because the integration is hard to replicate. @emollick's frame was simpler: the horse provides the power, the harness is what makes it useful. Two of the most-read technology thinkers, writing independently three months apart, describing the same structural shift. What does it mean when multiple people are converging on the same architectural conclusion? It means the market is figuring out what the durable bet is in the AI era. Not which model wins. Not which app gets the most users. The layer that makes the model accountable, routable, and governable inside a real organization. That is what Chiri has been building for two years.

Most AI implementations start with the tool. We start with the workflow. Before Chiri builds anything, we map how your team actually works. A few things about why that matters: The automation is only as good as the process it's automating. If the process is broken, faster just means broken faster. We capture the edge cases. The "actually, Sarah always checks this before it goes out" moments that make up 40% of the real work. That map becomes the ontology. The knowledge your team runs on, written down in a system the business owns. When the embed is done, you own it. You scale it. No building your home on rented land. Feels like the difference between installing software and actually changing how work happens.

With Chiri, humans stay inside the loop.

Giving everyone an AI license is a false congratulations. You made a procurement decision. Your credit card cleared. Congratulations! That is step zero. You still have to train them, govern the data, manage the cost, and prove the ROI. Buying AI is procurement. Getting ROI is operational. Don't confuse the two.

Turns out, your agents need headcount planning too.

Would you bet on one horse? 🐎 So why are you betting your company on one AI model?

Can you remember when all the code you wrote was free-range, organic, and grass-fed? 🌱 Our CTO breaks down the differences between organic and agentic engineering 👇

Be honest... what data have you been stuffing in your company's closet?

Your vibe coding is spaghetti 🍝

Alex Karp on CNBC: why is anyone paying for tokens that create no value? (July 2026) 73% of enterprises say AI costs beat their projections (FinOps Foundation, 2026). The value was never the token. It's the judgment and the routing wrapped around it.

We're burning billions of dollars a day running our smartest AI models on our dumbest tasks. As AI scales, the biggest lever on cost and performance won't be picking one "best" model. It'll be routing each task to the model that actually fits it. This piece by @tomas_hk breaks down why intelligent model routing is shaping up to be the next battleground for efficiency and economic value in AI, and what that means for how companies build and spend from here.

Great day to be model agnostic...

Token prices fell 67% last year. Enterprise AI budgets grew 483% in the same period. Five signals from the last 30 days explain why. A thread. 1. @OpenAI and @AnthropicAI both launched forward-deployed engineering businesses in May. $4B +. $1.5B JV. 224 open FDE roles across 39 companies. The labs are building what operators already figured out. 2. The spread between cheapest and most expensive production models today: 4,500x. Smart routing = $2.31/M tokens. Default routing = $18.40/M. Bills went up while prices dropped. The math problem is a routing problem. 3. @Gartner_inc, May 26: “Applying uniform governance across AI agents will lead to enterprise AI agent failure.” By 2030, 50% of deployment failures trace back to insufficient runtime enforcement. 4. @BessemerVP is now using the phrase “software-as-a-service turning into service-as-a-software.” SaaS multiples hit decade lows. AI gross margins: 50-60%. Mature SaaS: 70-90%. The re-rating is already a line item. 5. @azeem this week: only 27% of executives say AI has met ROI expectations, three years post-ChatGPT. Uber’s COO put it plainly: “one plus one plus one plus one equals one-and-a-half.” The gains pile up waiting for approval chains. Five signals, one story. As intelligence gets cheaper, value migrates to the layer that governs it, routes it, deploys it. The AI model is the raw material. The governance layer is the durable value. If your team is in the gap between a capable model and a measurable result, we’d be glad to talk about closing it. chiri.ai

What's keeping CTOs up at night in the era of AI? Our CTO Matthew Wimberly shares the good, the bad, and the ugly of AI governance and security.

AI governance moves like Treebeard. Your agents move at lightning pace. Dario Amodei wants national policy to catch up. The same gap sits inside your company: agents running ahead of any rules. You can close that one this quarter.

The cheap AI era is over. Copilot bills jumped 25 to 60x the day usage pricing landed. One team went from $50 a month to $3,000. The flat subscription was always a subsidy. Now you pay what it actually costs. The paradox nobody budgeted for: the price per token fell from $10 to $2.50 in a year, and your AI bill still went up. Tokens got cheaper. You just bought 24x more of them. Microsoft ran the numbers and pulled its own engineers off Claude Code. Roughly $2,000 per engineer per month in tokens versus $39 a seat flat. When the bill gets real, even believers route around it.

“𝗩𝗶𝗯𝗲 𝗰𝗼𝗱𝗶𝗻𝗴 𝗶𝘀 𝗮 𝘀𝘆𝗺𝗽𝘁𝗼𝗺, 𝗻𝗼𝘁 𝗮 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆.” Vibe coding did not appear because non-engineers suddenly became engineers. It appeared because two things were broken: engineers were too slow, and engineers were too far from the user to build what was actually needed. Spinning the work out to non-engineers fixes those two problems and breaks the two things engineers are genuinely good at: reliability and scale. 𝗪𝗵𝗮𝘁 𝗶𝘁 𝗳𝗶𝘅𝗲𝘀: speed, and fit to real user needs. 𝗪𝗵𝗮𝘁 𝗶𝘁 𝗯𝗿𝗲𝗮𝗸𝘀: reliability and scale, the exact disciplines that justify an engineering function in the first place. So it is a band-aid over an organizational design problem, not a cure for it. 𝗧𝗵𝗲 𝗲𝘃𝗶𝗱𝗲𝗻𝗰𝗲. Studies of AI-generated code find 40% to 62% carries security flaws, technical debt accrues roughly three times faster, code duplication climbs about 48% while refactoring drops about 60%, and practitioners report it is rarely safe for production, complex back-ends, or anything that has to scale. The pattern is consistent: fast and user-shaped, but fragile under load. 𝗖𝗵𝗶𝗿𝗶’𝘀 𝗽𝗼𝘀𝗶𝘁𝗶𝗼𝗻. Our pod operating model sits firmly in the middle: the speed and user empathy that make vibe coding attractive, paired with the reliability and scale that make it safe to keep. We bridge the gap instead of picking a side.