Yesterday we chased ByteDance UniVR-34B Planning and NVIDIA nvDock — open visual planners and science docking weights. Launch week and capital week are scored. Today’s delta is plumbing.

Who owns how agents connect to enterprise data, tools, and each other is no longer a side quest. Google, Microsoft, Salesforce, Snowflake, and ServiceNow are lining up a shared backend protocol explicitly framed as beating back Anthropic and OpenAI. Meanwhile Mistral shipped an open proof agent that does not “probably pass tests” — it checks Lean 4 math on real code and found bugs fuzzers miss. And OpenAI’s full-duplex GPT-Live is the consumer surface racing for ears before Gemini’s mid-July window.

If you ship coding agents into anything that touches CRM, cloud, or safety-critical paths, this is the Tuesday brief.

Protocol wars sound boring until they own the decade

According to The Information via weekend recaps, the enterprise software stack is coalescing around a rival agent backend protocol. The signatory list is the story: Salesforce, Snowflake, and ServiceNow sit on most of the world’s workflow and data gravity; Google and Microsoft own the clouds it runs on. The stated goal is not “interoperate with Anthropic’s MCP.” It is to stop Anthropic and OpenAI from owning the default agent connection layer.

That is a rational panic. Anthropic’s Model Context Protocol became the de facto tool-connection standard over the last year-plus. MCP is model-agnostic JSON-RPC plumbing: read files, call tools, push context. Major providers adopted it. Monthly download and call volume stories are now in the tens of millions. Once every IDE plugin, CRM connector, and internal RAG box speaks MCP, switching costs get ugly for anyone trying to sell a closed agent stack.

The irony layer: Microsoft, Google, OpenAI, and Anthropic are all platinum members of the Linux Foundation’s Agentic AI Foundation (AAIF) — the same foundation that took MCP as a founding contribution alongside OpenAI’s AGENTS.md and Block’s goose. Everyone cooperates in the foundation and knife-fights in the market. That is not hypocrisy; that is how standards wars always work. TCP/IP and HTTP looked “boring” until they decided who owned the internet.

Practical read for agent builders: do not rip out MCP tomorrow. Do not bet the farm that a five-company committee ships a coherent, fast alternative either. Keep adapters thin. Treat protocol choice as a router config, not a rewrite. If your harness hardcodes one vendor’s tool schema, you just lost the next standards cycle.

# Mental model, not production code
agent_tool_plane = [
  "mcp://legacy-and-indie",   # head start, open servers
  "enterprise://committee",   # Salesforce/MS/Google gravity
  "custom://your-monolith"    # always exists, always last mile
]
# Score: latency, auth, audit, who owns the connector when legal asks

If you are still wiring multi-agent dispatch by hand, remember last week’s Ollama capital + Muse Spark story: local and open stacks are funded escape hatches when protocol politics get expensive.

Leanstral 1.5: demand proof, not vibes

Mistral’s Leanstral 1.5 is the open-source counterweight to “AI wrote the PR, AI wrote the tests, ship it.”

• Apache-2.0, 119B total / 6B active (sparse MoE style efficiency story)
• Built for Lean 4 proof engineering and code verification
• Saturates miniF2F; 587/672 PutnamBench; SOTA open on FATE-H (87%) and FATE-X (34%)
• Multiturn RL loop: submit proof → Lean compiler feedback → refine until compile or budget exhausts
• Across 57 real repos, pipeline flagged 47 violated properties; 11 genuine bugs; 5 previously unreported on GitHub

The signature case study is brutal and useful: zigzag sign decode in datrs/varinteger overflowed on Std.U64.MAX — crash in debug, silent corruption in release. Tests and fuzzing often miss that class. Leanstral’s pipeline did not.

Why this matters for agentic coding: the industry is drowning in AI-authored code with AI-authored tests that share the same blind spots. Formal verification will not replace your unit suite. It is the layer for crypto, parsers, concurrency invariants, and safety-critical kernels where “probably works” is a liability. Pair a SWE agent that generates with a proof agent that refuses to greenlight without compile. That is a better architecture than another leaderboard screenshot.

# From Mistral's docs path
uv tool install mistral-vibe
vibe --setup
# /leanstall inside vibe, then:
vibe --agent lean
# Optional: Lean LSP MCP for editor-grade feedback loops

Weights: mistralai/Leanstral-1.5-119B-A6B. Free API endpoint as leanstral-1-5. If you already run local MoE serving on a used high-VRAM GPU, this is a different job class than chat — treat it as a verifier sidecar, not a daily driver coding model.

GPT-Live: full duplex, then delegate

OpenAI’s GPT-Live (shipped July 8, still rolling through the ecosystem this week) is not a coding agent. It is the consumer interface bet: full-duplex voice that listens and speaks continuously instead of walkie-talkie turns. Backchannels (“mhmm”), interruptibility, live translation, and delegation to frontier text models (GPT-5.5-class and successors) for search and deeper work while the conversation keeps flowing.

Architecturally that is the same idea as multi-agent stacks: a low-latency interaction plane + a heavy reasoning plane. Voice just makes the latency budget unforgiving. Paid tiers get the larger GPT-Live-1; mini becomes the default ChatGPT Voice path for the mass market OpenAI claims is already >150M weekly voice/dictation users.

For builders: if your agent product only has a chat box, voice is the next UX war Google cannot win with an API price cut alone. If your agent product has no tool protocol strategy, the enterprise alliance will decide your connectors for you.

What not to rehash

Sol / Terra / Luna / Grok 4.5 launch day is scored. Coding Agent Index cost matrix still holds. Langflow CISA KEV and EDR false positives are last week’s security homework. UniVR visual planning is yesterday’s open multimodal lead. Today is standards + verification + voice interface.

Gemini 3.5 Pro mid-July countdown (~July 17 in tracker chatter) still has no settled public GA card in your console. Leave a seat in the eval matrix. Do not re-architect on rumor dates.

Stack notes for Tuesday

1. Inventory MCP surface area. List every agent tool that speaks MCP. Document auth, scopes, and who operates the server. You will need that map when enterprise IT asks for the “other” protocol.
2. Add a verifier lane. Leanstral-class models will not fix your flaky integration tests. They will catch overflow and invariant bugs your agents invent. Wire as CI sidecar, not chat toy.
3. Keep protocol adapters thin. Hardcoded vendor schemas are tech debt with a political fuse.
4. Voice is product, not just demos. If customers talk to agents, full-duplex expectations are now the bar — even if your backend is still text tools.
5. Security still applies. Protocol connectors that store long-lived API keys in flow JSON are the same class of risk as last week’s Langflow story. Rotate. Scope. Prefer short-lived tokens.

Useful desk kit if you dual-run cloud coding agents next to local verifiers: a solid USB-C docking station for multi-laptop farms, a mechanical keyboard that survives overnight harness runs, enough external NVMe storage that 100B+ sparse checkpoints do not eat the OS disk, and a quiet USB condenser mic if you are actually dogfooding full-duplex voice agents instead of demoing them once.

The frontier coding war is about who owns the IDE default. The protocol war is about who owns the pipe under every agent. The proof war is about whether AI-written software can be trusted past the test suite. MCP has the head start. The enterprise alliance has the customer list. Leanstral just made “ship the PR” a little less of a vibe. Score all three before Gemini’s calendar alarm goes off.

Affiliate disclosure: Amazon links above use our Associates tag (spaghettistor-20). We only link tools we would actually buy for agent and local-model work.