Welcome back technologists🫡

A trade deal between the US and Canada was reached earlier in the week, put on hold while both sides finalized details, and then collapsed less than an hour before a midnight deadline. (about 4 minutes)

Here's the signal today: 🛸
👨‍💻🐕️

⚠️ Canada-US deal collapses, 50% tariffs land on $28 billion
🔌 Microsoft's policy plane cut agent spend 78%
🔌 Route prompts, save two thirds, keep quality
🔌 A trace judge for one hundredth the price
🔌 Ollama tests Claude model management
🔌 AWS asks who gave your agent the database password
🛡️ OpenAI copyright MDL gets an amended complaint
🛡️ Judge voids visa ban covering 40% of the world's nations

⚠️ The Friday Deadline

The deal that died before midnight. A tentative Canada-US trade agreement collapsed less than an hour before a late Friday deadline on August 21, triggering 50% US tariffs on $28 billion of Canadian exports at 12:01 a.m. Saturday. Prime Minister Mark Carney said the US tried to change the deal at the last minute, so he broke off talks, recalled his negotiators, and vowed "dollar-for-dollar" retaliation. Canada had spent a year seeking relief from existing levies: 50% on steel and aluminum, 25% on autos, 10 to 25% on forestry products. Both sides now escalate instead. If you sell physical goods, hardware, or anything with a cross-border supply chain, your input costs just moved. If you sell pure software, you are watching a live lesson in why your customers' budgets get cautious when tariffs make headlines. 💰 Money play: if any supplier or fulfillment step touches Canada or US steel, aluminum, or autos, reprice this week, not next quarter. Retaliation rounds compound, and the first mover on pricing keeps the margin.

🔌 Agent Watch

Who spent all the tokens. Microsoft engineers Tisha Chawla and Susheem Koul showed a FinOps policy control plane for agents that cut costs 78% while raising task completion rates from 67% to 96%. Read that pair again: cheaper AND more completions, because the plane kills wasteful retries instead of killing runs. If your agent bill is a mystery line item, this is the talk to watch before your next invoice. 💰 Money play: put a per-run token budget in front of your agent loop this week; the Microsoft numbers say governance pays for itself.

Preferences over benchmarks. DigitalOcean's Archana Kamath and Tyler Gillam made the case for routing prompts to the right model instead of defaulting to one premium model, cutting costs by two thirds while holding quality. The leaderboard winner is a pricing trap for most of your traffic. 💰 Money play: classify your top five prompt types, point the cheap ones at a cheap model, and keep the frontier model for the hard tail only.

A judge for a penny. LangChain processes billions of tokens a day across production traces, so it partnered with Fireworks to fine-tune a Qwen model that detects "perceived error," the moments a user thinks the agent messed up, on every trace. It matched or beat frontier-model judging and runs up to 100x cheaper. Evals stop being a sampling exercise at that price. 💰 Money play: if you run agents in production, score every trace for perceived error instead of spot-checking; LangSmith has an early tester signup on the post.

Ollama meets Claude. Ollama tagged v0.33.0-rc0 on August 21 with one line that matters: app support for Claude model management. It is a release candidate, so treat it as a preview, but the local-model app is reaching for the frontier APIs. Worth a test install if you live in Ollama already.

Who granted that access. AWS opens its AgentCore Gateway post with a scene that should make you sweat: a teammate's laptop, an mcp.json file, a production database password in plain text, and a comment reading "TODO: rotate this." The post walks through centralizing and auditing which tools MCP-enabled assistants like Claude Code, Cursor, and Kiro can touch. 💰 Money play: grep your machine and your repo for mcp.json today, and rotate anything that answers back.

🛡️ The Shield

The scraped-data case grinds on. Entry #1677 hit the docket in In Re: OpenAI, Inc. Copyright Infringement Litigation, the MDL (1:25-md-03143) in the Southern District of New York, with an amended complaint continuing claims over training on scraped works. The docket was last updated August 21. This is the case family that could decide whether builders may freely train on scraped data, so every fine-tune you ship rides on it. 💰 Money play: keep provenance notes on your training and fine-tuning data now; "we will figure out licensing later" is a strategy with a court date.

The visa ban falls. Judge Jeannette Vargas in the SDNY struck down Secretary of State Marco Rubio's January order denying immigrant visas to citizens of 75 countries, about 40% of the world's nations. Vargas called the order "patently unlawful," ruling that the Immigration and Nationality Act explicitly strips the Secretary of authority over consular adjudication of immigrant visas and requires individualized assessments of each applicant. Catholic Legal Immigration Network and African Communities Together brought the suit. For nomad founders and anyone hiring remote talent across Latin America, Africa, Southeast Asia, the Middle East, or the Caribbean, a real door reopened on August 21. 💰 Money play: if a hire or a co-founder was frozen by the ban, have them talk to an immigration lawyer this week; struck-down orders still need consulates to act, and the early filers get the early interview slots.

Sponsored by

Smarter CRM. Less Busywork.

Disconnected data and tools make it harder to understand your customers. HubSpot's Agentic Customer Platform brings your data, teams, and tech stack together with AI built in to help your business work faster and create more personalized customer experiences.

Why HubSpot and what's new

• Use AI powered tools to take action faster

• Unify your data, teams, and tech stack in one place

• Create one shared view of customer data

• Connect teams around the same customer context

• Bring your business tools into one place

Connect more of your business in one place and give every team a smarter way to work. Get set up quickly and start checking off your hardest tasks.

Get started free

🎓 Level Up

Level Up: the 20-minute agent cost audit

The Microsoft and DigitalOcean talks point at the same Monday verb. Run this against your most expensive agent today:

AGENT COST AUDIT (20 minutes)

1. EXPORT. Pull the last 500 runs from your trace
   store (LangSmith, Helicone, or raw logs).

2. BUCKET. Tag each run with one of your top 5
   prompt types. If you cannot name your top 5,
   that is finding number one.

3. PRICE. For each bucket: total tokens, cost,
   completion rate, retry count.

4. ROUTE. Any bucket above 60% of runs that a
   mid-tier model handles: repoint it. Keep the
   frontier model for the hard tail only.

5. CAP. Set a per-run token budget at 2x the p90
   of successful runs. Kill and log anything past
   it instead of letting retries burn.

6. JUDGE. Score every run for "user thought this
   was wrong," not just sampled ones. A fine-tuned
   small model does this for pennies (see the
   LangChain trace-judge post above).

Target: 40%+ cost cut in week one. Microsoft's
policy plane got 78% with HIGHER completions, so
the ceiling is nowhere near your floor.

Reply and tell me what your audit found in step 2. I read every answer, usually with Kobe supervising.

If this issue earned its five minutes, forward it to one builder who is still routing everything to the expensive model. They can subscribe at nomadsignal.ai.

See you tomorrow,
Chase & Kobe 👨‍💻🐕

One tap. Feedback goes straight to the desk.