|
Hey there 👋
Two things happened at once this week, and they point in opposite directions. Agents kept quietly shipping into real production — regulated pharma R&D, live catalogs, your Copilot seats, the security perimeter. And the biggest names in enterprise software started openly fighting over the plumbing those agents run on.
That second one is the story. For a month we’ve said the barrier was never the model — it was the boring work of wiring an agent into your systems and trusting it there. This week the fight moved one layer deeper: to whose connection standard everyone wires into. Your instinct on reading that will be to pick the winner. Don’t. The move today is the opposite, and it’s cheaper.
The Big Thing
Five of the biggest names in enterprise just lined up against Anthropic’s MCP
Here’s what got reported Monday, and why it matters more than any single launch. According to The Information, five of the heaviest hitters in enterprise software — Google, Microsoft, Salesforce, Snowflake, and ServiceNow — have agreed to co-support a shared “backend” approach for how AI agents connect to enterprise data and tools. The framing is explicit: a counter to Anthropic and OpenAI on the agent-infrastructure layer.
The thing they’re circling is MCP — Anthropic’s Model Context Protocol, the way agents plug into tools and data. For about 18 months it’s been the de facto standard, and Anthropic hardened that lead by donating MCP to the Linux Foundation’s Agentic AI Foundation back in December. When it’s the default under OpenAI, Google, Microsoft, Amazon, and Salesforce alike, “de facto standard” undersells it.
Now the third-thought read, because “protocol war” is the wrong frame. Those same five incumbents already support MCP today — none of them is ripping it out. And A2A, the agent-to-agent protocol, was designed to complement MCP, not replace it: MCP connects an agent to tools, A2A connects agents to each other. So this isn’t a winner-take-all cage match. It’s the giants moving to own the orchestration layer that sits above the tool layer MCP already won. The real tell? Five companies this size aligning against Anthropic’s standard is the loudest possible signal of how completely that standard landed.
Ship it? Watch — and do not re-architect anything. There’s no spec, no product, and no ship date here; this is reporting on strategic alignment, not a launch. But it hands you one cheap move this week: buy portability insurance. Keep your tool access MCP-shaped. Keep your harness and your model swappable. Don’t hard-wire your integration layer to any single vendor’s orchestration protocol, no matter how much of your stack they already run. Whoever wins the standard gets default status in every deployment for years. Whoever stays abstracted keeps their leverage the day the default changes.
Sources: The Information (Jul 13, reported — paywalled); corroborated across independent outlets. MCP → Linux Foundation Agentic AI Foundation, Dec 2025. This is positioning, not a shipped spec.
Tour de Headlines
🧩 An open-weight coding agent just landed in seats you already pay for. On Jul 7, GitHub made Kimi K2.7 Code selectable in Copilot’s model picker for Business and Enterprise — the first open-weight model in the lineup. It’s a trillion-parameter mixture-of-experts (~32B active), Azure-hosted, billed by usage. One catch worth respecting: it’s an admin opt-in, off by default, so review governance before you flip it. The play is small and concrete — turn it on for one team and A/B it against your default model on real pull requests. And it’s quietly on-theme with today’s Big Thing: an open-weight model is a swappable model. That’s portability you can act on this afternoon.
🛡️ A control plane that watches every hop an agent makes. Also this week, First Recon AI shipped its AI Security Runtime (GA, Jul 8). It inspects every AI interaction — human-to-model, agent-to-tool, and agent-to-agent — and applies policy inline, before data reaches the model. The part that matters for anyone who has to answer to an auditor: it records each decision as sealed, audit-ready evidence mapped to NIST, GDPR, and the EU AI Act, and it feeds your SIEM. Adopt agents with a governance trail from day one instead of bolting it on after the incident. There’s a 30-day trial. File it under the moat a standards war can’t commoditize.
🧬 The delight: an agent that reasons over models a regulator already trusts. On Jul 7, Certara and NVIDIA paired up on the BioNeMo Agent Toolkit — turning an agent into an “autonomous life-sciences scientist” that reasons over validated biosimulation models and datasets, not just text. Think dosing optimization, clinical-dataset interrogation, trial simulation, ADMET prediction — with a scientist kept in the loop and regulatory-ready evidence at the end. Certara serves 2,600+ biopharma, academic and regulator clients, so this isn’t a demo reel. The tell for the rest of us: the agents landing in the hardest, most-regulated rooms win by grounding themselves in domain models someone already certified — not by being chattier.
|
Sponsor
The protocol war won’t fix your sales calls.
While the giants fight over agent plumbing, the highest-stakes workflow in your company still runs unmeasured — your live sales and CS conversations. The misread discovery call. The deal where two people “aligned” on different things. The rapport that never formed, and closed-lost with no warning light. RapportScore reads the human signals in every call and email and scores how your team actually connects — B2B communication intelligence for teams that live on calls. You’re verifying the agents. Verify the conversations too.
See your team’s score →
|
Tool of the Day
🚣 rowboat
A local-first, open-source AI coworker that builds a plain-Markdown knowledge graph from your notes — and acts on it.
If the Big Thing is about staying portable, rowboat is what portable looks like in your terminal. It’s an open-source (Apache-2.0, ~15k GitHub stars) AI coworker that runs on your machine: it reads your email and meeting notes, builds a knowledge graph as plain, inspectable Markdown, and acts on it. Tools come through MCP and Composio; the model is bring-your-own — Ollama, LM Studio, or a hosted API, your call. Nothing about it is locked to a vendor’s orchestration layer, which is the whole point this week. Deploy it on exactly one workflow — meeting prep, or drafting an email that actually remembers your last three threads with the person — and keep the data as files you can read. Free, no purchase decision, and every piece of it is swappable.
Clone the repo →
Worth a Click
- Prove your code is right, not just tested — Mistral shipped Leanstral 1.5 (Apache-2.0, free API and open weights), an agent for code verification, not just generation. Mistral says it turned up 5 previously-unreported bugs across 57 real repositories, including a live overflow bug. Point it at one critical function and have it prove correctness properties your AI-written tests would sail right past. Verification is exactly the kind of moat the standards war can’t touch.
- Your next support agent will listen and talk at the same time — OpenAI shipped GPT-Live (Jul 8): full-duplex voice models that listen and speak simultaneously — they’ll even drop an “mhmm” while you think — and hand off to a frontier model in the background for the hard reasoning. It’s rolling out in ChatGPT now. The reason to file it: this is the voice substrate the next wave of support and CS agents get built on — scope your roadmap around it.
- The map behind the Big Thing — before you believe anyone’s “protocol war” headline, read how the agent standards actually layer — MCP for agent-to-tools, A2A for agent-to-agent, WebMCP and the rest stacking rather than fighting to the death. It’s the clearest picture of why “stay portable” beats “pick a winner.”
Step back and the week tells one story from two angles. Agents shipped into the hard rooms — Certara into regulated pharma, Leanstral into your codebase, Kimi into your Copilot seats, First Recon onto the wire — while the giants started fighting over the plumbing all of it runs on. Trace the arc: the barrier was never the model, then it was the integration labor and the trust, and now it’s whose connection standard wins. Here’s the move that survives whoever wins: don’t pick a side. Stay portable — MCP-shaped tools, a swappable harness and model — and pour your effort into what no standards war can commoditize: governance, verification, audit-ready evidence. Whoever owns the standard gets default status. Whoever stays abstracted keeps the leverage.
Build like the default will change — because it might.
— Ace, for The Agent Stack
|
