Forget the megacap earnings scoreboard for a day and look one layer down, at the plumbing you're actually going to build on. This week the agent stack moved on three fronts at once: it got cheaper to run, more standardized to build against, and demonstrably easier to break into. Those three facts belong in the same post because they're the same story — the layer where AI stops being a demo and starts being production is repricing and re-exposing itself in real time, and you're the one who has to decide how fast to lean in.
OpenAI makes inference cheap enough to stop thinking about
OpenAI cut API prices on two of its lower-cost models on July 30, and the size of the cut is the story. The cheapest tier dropped 80% and the mid-tier fell 20%, while the flagship price held: the budget GPT-5.6 Luna now runs $0.20 per million input tokens and $1.20 per million output, down from $1 and $6; the mid-tier Terra fell to $2 / $12 from $2.50 / $15; and the flagship Sol keeps its $5 / $30. The framing from OpenAI's side was explicit — it's cutting prices as companies grow more sensitive to costs, which is a polite way of saying the cheap open-weight models have been eating its lunch all month.
The operator's take: an 80% cut doesn't just lower a line item, it changes what you're allowed to build. Workloads that were too expensive to run at scale — classify every ticket, summarize every call, screen every document — cross into "just do it" territory at 20 cents a million tokens. Stop budgeting by cost-per-token and start budgeting by cost-per-task, because the token price is now falling fast enough that it's the wrong unit to plan around.
The agent plumbing just standardized under you
While pricing grabbed the headlines, the protocol layer made a bigger structural move. The Agentic AI Foundation, a Linux Foundation directed fund, shipped the July 28 Model Context Protocol spec — the largest change since launch — and it makes MCP fully stateless, dropping sessions entirely. That sounds like a footnote until you remember MCP is becoming the USB port between your models and your tools, and a stateless design is the one that actually scales behind a load balancer without sticky-session gymnastics.
The operator's take: protocol standardization is the quiet part that decides whether agents graduate from pilot to platform. Stateless is the right call for anyone running this in production — it's the difference between an architecture your infra team recognizes and one they fight. If your teams are building against MCP, this is a version bump worth reading before it's a migration you're forced into, and worth putting on the roadmap now rather than in the incident review later.
The security bill for the agent rush came due
Then the other shoe. Noma Labs disclosed CVE-2026-59726, nicknamed "RufRoot" — a maximum-severity CVSS 10.0 flaw in the open-source agent platform Ruflo, which has 67,000-plus GitHub stars — where a single unauthenticated HTTP POST to the MCP bridge on port 3001 could execute code and exfiltrate LLM API keys. Unauthenticated, remote, code execution, and it walks out with your model keys: that's about as bad as a score gets, in a project that a lot of teams pulled in precisely because it was popular and free.
The operator's take: the same cheap, standardized agent stack that makes the first two stories exciting is the one that just shipped a 10.0. Every MCP bridge you stand up is a new listening socket with reach into your models, your data, and your credentials — treat it like the internet-facing service it is, not like a dev convenience. Inventory what agent frameworks are already in your environment, because "we grabbed the popular open-source one" is how a starred repo becomes your breach.
Also on my radar
- Coinbase disappointed on Q2. Shares fell after the exchange posted weaker-than-expected second-quarter results. Crypto-linked revenue is still a trading-volume business dressed up as a platform — a useful reminder when a vendor's "recurring" line is really a market bet.
- Reddit dropped 11% on shaky referral traffic. The stock sank even as results beat estimates, on "choppy" search referrals. Anyone whose funnel leans on Google-sent traffic should read that as a warning that the front door keeps moving.
- Memory prices are the new capex tax. Korea's chip complex ripped this week — the KOSPI closed July 31 at 6,595.45, up 17.9% in its largest single-day gain ever, with SK Hynix limit-up and Samsung up 27%. The same memory demand that has Amazon citing higher costs will show up in your hardware and cloud quotes too.
The throughline for the day: the layer where AI becomes real work just got cheaper to run, cleaner to build on, and provably easier to attack — all at once. The cost of adopting agents dropped this week; the cost of adopting them carelessly went up by exactly the same measure. Plan for both. That's the Signal for today.
Paul Sapio is the CIO of Mikhail Education and a full-stack AI engineer. Open to contract work in security, networking, AI, and SaaS development — reach out.