AI Digest, Oct 12: Autonomous Agents, Tariff Inflation, Verifiable AI

0:02PM Lead ProThis is PM Lead Pro. Today's digest covers autonomous agents, tariff-driven inflation and verifiable AI, plus a quick model radar, with takeaways for teams adopting AI.

0:17You're probably commuting right now, you know, just watching the world go by.

0:21But your newest colleague, they are already at their desk.

0:25Right, they're clicking through web browsers, reading files.

0:29Yeah, and directly messaging coworkers to close out tickets.

0:32Except they aren't actually human.

0:34No, definitely not.

0:36I mean, we are officially crossing that threshold from passive chatbots to autonomous employees.

0:43Anthropic just dropped their computer use skills and files APIs into general availability.

0:47And Google is right there with them, right?

0:49Collapsing their enterprise AI into a single Gemini agent locked inside this new agent sandbox.

0:55Exactly.

0:56But what really caught my eye in the agent report is the new Hermes agent, version 0.162.0.

1:01They just introduced peer-to-peer bot DMs.

1:05Wait, like they can message each other?

1:06Yeah.

1:07An AI agent can now silently message another bot to hand off a task, which completely shifts the threat landscape.

1:15Oh, wow.

1:15I mean, if a Hermes bot can privately DM another bot, who is actually auditing that conversation?

1:23Exactly the problem.

1:24The traditional API limits you set up, they don't mean much if your agent gets, you know, socially engineered by an external bot in a private channel.

1:33So are Google's new sandboxes genuinely secure or are they just like digital playpens?

1:38Well, Nate Soares was on the Odd Lots podcast recently, and he basically compared AI labs to alchemists breathing mercury.

1:46Right, while claiming they're perfectly safe just because they installed a fume hood.

1:50Yeah it a perfect analogy The mercury here is the raw unaligned model weights and the fume hood is that external sandbox So the hood might vent the obvious smoke but it doesn stop the invisible systemic leaks

2:04Exactly. Especially when you have people inside the lab actively trying to, you know, bottle that mercury.

2:10PM Lead ProAgents can now message each other privately, and standard API limits may not stop an agent being socially engineered by an external bot. Before deploying, ask who audits agent-to-agent conversations.

2:25Oh, you mean the South China Morning Post report?

2:28Yeah. Anthropic researcher Jacob Coxon just resigned, stating he is 90% certain Chinese spies are currently inside top Western AI labs.

2:3790%. I mean, when you have autonomous agents writing code, combined with the constant threat

2:42of state-sponsored espionage trying to insert backdoors, you start to see why the U.S. is

2:47aggressively trying to physically isolate its supply chain. Which brings us directly to the

2:52tariffs. Right. Because this geopolitical friction, it's instantly hitting the bottom line for any

2:57founder trying to secure compute. Yeah, new research from the CEPR lays out the 2025 to

3:02A one-point tariff increase actually pushes up consumer prices by 0.25% after a year.

3:11And the numbers are wild. By February 2026, tariffs added a peak 2.9 percentage points to U.S. goods inflation.

3:18But the mechanism behind that inflation is what engineering leads really need to understand.

3:23Only two-thirds of that 2.9% comes from the direct cost of imports.

3:28Wait, where does the other third come from?

3:29Domestic manufacturers quietly jacking up their own prices.

3:34Hang on, if the domestic producers are supposed to be protected by the tariff, why are they raising prices?

3:40Is that just price gouging under the radar?

3:43Well, partially.

3:44But it's mostly because local manufacturing still relies super heavily on imported raw materials.

3:50Oh I see Yeah when those imported inputs cost more domestic producers just pass that cost to you So if you forecasting your AWS bill or buying local hardware this year

4:00those protective tariffs are essentially acting as a backdoor tax on your runway.

4:04Exactly. Meanwhile, China is taking a totally different approach to both the hardware

4:09constraints and the safety problems. Right, SCMP notes they are aggressively

4:13building out molten salt and supercritical CO2 infrastructure.

4:17And before we glaze over those terms, this isn't just random chemistry. It's advanced thermal energy storage.

4:24So they use these extreme heat systems to store intermittent renewable energy and then discharge it as a stable, continuous baseload.

4:31Precisely. Because frontier AI data centers cannot run on intermittent solar power.

4:36They need 24-7 grid stability, and China is physically hardwiring that into their grid.

4:41Wow. But their software approach is even more interesting.

4:44Didn't Tsinghua University just open source Veriloop E2?

4:48Yeah, they are calling it verifiable intelligence.

4:51How does that actually work?

4:52Because if a Hermes bot is hallucinating in a private DM, like we said, an external sandbox won't necessarily catch it.

4:58Right, so instead of relying on a sandbox to catch bad behavior after the fact,

5:03Veriloop E2 mathematically checks its own long-chain reasoning during generation.

5:07Oh, like formal verification in software engineering.

5:10Exactly. If the model encounters a logical contradiction in its math, the loop just halts. It forces a recalculation before it even outputs the token.

5:19That seems like a massive paradigm shift. You embed the safety into the mathematical logic of the model itself, rather than wrapping a dangerous model in a flimsy perimeter.

5:28Which brings us to the practical takeaway for anyone deploying these autonomous systems today Do not rely on generic API limits Right you need strict authorization perimeters at the database level Yes implement cryptographic identity and OAuth gating

5:43If a rogue Hermes bot gets tricked in a DM,

5:46that bot's specific cryptographic signature

5:48should physically prevent it

5:50from making unauthorized changes.

5:51Great advice.

5:53PM Lead ProThe practical advice here: don't rely on generic API limits. Use authorization perimeters at the database level, with cryptographic identity and OAuth gating, so a tricked agent cannot make unauthorized changes.

6:07Okay, let's hit the model radar real quick.

6:10What are the top new releases moving the needle this week?

6:13Well, DeepMind just dropped Gemma 4.

6:15It's a 12 billion parameter multimodal model, but the real big deal is that it's encoder free.

6:21Dropping the encoder layer massively reduces the memory footprint, right?

6:25Yeah. So you can now run frontier level intelligence locally on a standard personal computer without crushing latency.

6:31That is huge. Next up is Qen3 Guard. This isn't a generalist model. It's built purely for real time safety classification.

6:38Right. It sits directly on live token streams and evaluates them instantly, which ties perfectly into that continuous verification approach.

6:47And if you need raw speed, artificial analysis reports DeepSeek v4.1 Flash is absolutely dominating.

6:54Oh, yeah. It's spitting out 208 tokens per second at an incredibly low cost per task.

7:00I mean, 208 tokens a second is practically unreadable for a human monitor, which leaves us with this final provocative thought.

7:08If autonomous agents can now privately DM each other at over 200 tokens a second and literally rewrite their own code,

7:16how long until they invent a highly compressed language to communicate with each other that human overseers cannot even decipher?

7:23Let's just hope those mathematical fume hoods are tightly sealed.

7:28Keep exploring and thanks for joining the deep dive.

7:32PM Lead ProMore at pmlead.pro, where the transcript and sources are on the episode page. A question for you: who audits what your agents say to each other?