Five verified moves this week, and a team of one can act on each before Monday. The most-used agent protocol dropped sessions in its biggest revision since launch. OpenAI cut its cheap tier by 80% — three weeks after shipping it. The largest open-weight model ever released landed on Hugging Face and immediately outranked a US frontier model on coding. Google's free agent curriculum is the thing your feed is sharing. And the money quietly moved toward governing what agents are allowed to do. Every item below is dated and sourced; each carries the one line that changes what you do next.
1. MCP's stateless spec went final — you can delete infrastructure#
The 2026-07-28 Model Context Protocol specification shipped July 28, and it's the largest revision since MCP launched (MCP blog). The headline is that sessions are gone: the spec removes the initialize/initialized handshake and the Mcp-Session-Id header. Each request now carries protocol version, client identity, and capabilities in _meta. It also adds header-based routing (Mcp-Method, Mcp-Name), cacheable tools/list responses with ttlMs, Multi Round-Trip Requests to replace held-open streams, and hardened OAuth. Tier-1 SDKs — TypeScript, Python, Go, C# — shipped stable support on day one.
What it means: A remote MCP server that used to need sticky sessions, a shared session store, and deep packet inspection at the gateway can now run behind a plain round-robin load balancer. That's real infrastructure you get to remove. Existing servers keep working, but new clients speak stateless by default — so claim the deploy simplicity now, while your traffic is small. We walked through shipping a stateless server on the stable SDK today.
2. OpenAI cut Luna 80% — the price floor is collapsing#
On July 30, OpenAI cut GPT-5.6 Luna by 80% — from $1/$6 to $0.20 per 1M input / $1.20 per 1M output — and Terra by 20%, to $2/$12. Flagship Sol is unchanged at $5/$30 (VentureBeat; CNBC). The cut lands three weeks after the family launched on July 9, and OpenAI framed it around cost-sensitive customers — reporting has tied the pressure to Chinese open-weight models capturing a growing share of US enterprise token usage.
What it means: Your agent's default tier is now mispriced in your favor. A high-volume extraction, classification, or drafting agent that ran on Luna at launch just got five times cheaper without a code change. Re-run your cost-per-completed-task math this week — and if you're on Sol for work that Luna can do, route by tier deliberately instead of defaulting to the top of the range.
3. Kimi K3's open weights landed — and beat Fable 5 on frontend code#
Moonshot released Kimi K3 open weights on July 26–27: a 2.8-trillion-parameter Mixture-of-Experts with ~104B parameters active per token and a ~1M-token context, shipped as roughly 1.56 TB across 96 shards on Hugging Face (Tom's Hardware). It's the largest open-weight model ever released, and it ranked #1 on the LMArena Frontend Code evaluation, ahead of Fable 5, while topping the Artificial Analysis open-weight intelligence index. Together AI and Modal offered day-0 hosted access; the API runs about $3/$15 per 1M tokens (OpenRouter).
What it means: The gap between "frontier" and "open weight" just narrowed to a benchmark rounding error on at least one real task. You don't have to self-host a 2.8T model to benefit — the existence of a credible open-weight option is what gives you leverage over any single vendor's pricing. Add one to your model router as a fallback now, before a price shock or a rate limit makes it urgent.
4. Google's free agent curriculum is the week's most-shared link#
Google and Kaggle's AI Agents Intensive is circulating hard this week, now available as a self-paced Kaggle Learn Guide (Kaggle). Built by Google researchers and engineers, it pairs a whitepaper with hands-on codelabs on Gemini and the Agent Development Kit across the topics that actually matter: agent memory, agentic loops for long-running agents, building an MCP server, and multi-agent systems. The first edition reached more than 1.5 million learners.
What it means: It's free, it's current, and it maps cleanly onto the gaps most solo builders have. Don't work it front-to-back — pick the one module that's a hole in your stack. If it's memory, pair it with how to read an agent-memory benchmark; if it's MCP, pair it with item 1 above.
5. The money moved to agent identity#
Hush Security added a $30M Series A on July 28 — Akamai joining Battery Ventures and YL Ventures, bringing its total to $41M — to govern non-human identities and the actions AI agents take (Tech Startups). It's one of several rounds this month funding the scoping, granting, and logging of what agents are allowed to do.
What it means: The moment your agent can spend money, delete data, or push code, its permissions are a security surface — and "the agent did it" is not an audit trail. You don't need to buy anything to get the lesson: give each agent a scoped credential and an action log now, the way you would a new hire. We made the same argument when YubiKey shipped verified authorization for agent actions.
The throughline: the cost of raw intelligence is collapsing (items 2 and 3) while the plumbing and the guardrails around agents professionalize (items 1 and 5) — and the skills to build on both are now free (item 4). For a founder, that's the whole job this week: re-price your defaults, add an open-weight fallback, delete the session infra you no longer need, and give your agents scoped credentials before someone else's incident makes you.
