Hey folks,

This week we got into the latest wave of model releases, but the one that caught our attention was Jev, a new kind of decision model built to make fast, structured judgements rather than hold a conversation.

We talked about what happens when classification, prioritisation and routing become cheap enough to happen almost everywhere. Instead of throwing a frontier model at every small decision inside a workflow, Jev can make those calls in milliseconds and at near-zero cost, leaving the larger models to handle the parts that actually require open-ended reasoning. That could change everything from business workflows to the way coding agents navigate repositories and choose tools.

Then we got into Meta’s launch of Muse and whether personal agents are about to have their mainstream moment. Meta already has distribution through products like WhatsApp and Instagram, which raises an interesting question about whether reach matters more than having the most capable model.

And if everyone does end up with an agent, we talked about what that does to businesses built around consumer inertia. Insurance, loans and subscriptions look a little different when an agent can keep comparing prices, renegotiating terms and trimming things you no longer need without waiting for you to get around to it.

Intelligence Snacks now has a new Graph view that lets you explore sideways through the various ideas, themes and concepts that come up on the pod, a bit like a jungle gym, rather than navigating each episode hierarchically. As more episodes are added, new connections form around the existing material, so older ideas can gain new context instead of simply disappearing further into the archive.

You can explore the graph and knowledge mix at intelligencesnacks.com/graph/.

A large share of a coding agent’s work can be repository navigation rather than writing code. It lists files, opens one, inspects the result and chooses the next tool. Each narrow choice can trigger another model completion carrying the conversation, system prompt and accumulated context. Prompt caching may soften that cost, but deciding which file to open still doesn’t require the same capability as designing or reviewing a code change.

An agent harness could split those jobs. A fast decision model such as Jev could select tools, route outputs and choose the next action, while the frontier model handles architecture, difficult reasoning and code generation. Because the decision layer is built for structured classification rather than conversation, it can return predictable JSON quickly and at negligible cost compared with repeatedly asking a frontier model to manage every step.

Code graphs could sharpen that division further. A graph can map relationships across a repository, trace the impact of a front-end function and identify the code relevant to a proposed change. A decision model could then help navigate that structure and narrow the context before the frontier model begins its substantive work. Instead of paying for broad intelligence to rediscover the repository on every tool call, the harness would apply it after cheaper components have found the right files and context.

Many automated workflows still send narrow judgement calls to frontier language models. A model may only need to classify an item, choose a route or check an output, yet each call carries the latency and cost of a far more capable system. Jev takes a different approach. It is designed to return structured decisions with confidence scores, making those small judgements in milliseconds.

That changes where intelligence can sit in a workflow. Andy used Jev to rank SEO signals before an agent investigated them, replacing confidence scores he suspected the agent was largely fabricating. Pete used it to classify which material retrieved from a company graph should enter an agent’s prompt. In both cases, Jev handled the sorting and prioritisation before the more expensive agent started its main task.

The economics matter as much as the speed. Jev’s charges were effectively invisible in Andy’s OpenRouter usage, while Pete found the model cheap enough to consider inserting it throughout longer chains of decisions. Instead of paying a frontier model to select tools, route items or make other repetitive judgements, builders can reserve larger models for open-ended reasoning and use specialised decision models for the repeated decisions around it.

Meta’s launch of Muse might be the moment personal agents start going mainstream. That becomes particularly interesting in industries like insurance, loans and subscriptions, which often benefit from consumer inertia. Comparing alternatives takes time, the savings remain uncertain until the search is complete, and switching can involve more effort than an unhappy customer wants to spend. Businesses can therefore keep earning from people who know they might get a better deal but never quite get around to looking for one.

Personal agents could absorb that repeated work. An agent could continually compare prices, renegotiate terms and cancel subscriptions that no longer provide value, without waiting for its owner to find the time or motivation. That changes more than convenience. Occasional bargain hunting becomes a continuous process, and a service priced for passive customers instead faces an automated buyer that keeps returning to the market.

The shift would also create a fight over who controls the path to a purchase. An agent becomes more valuable when it can search and act across shops and services, while commerce platforms have reason to keep transactions inside the experience they control. Amazon has already blocked Muse from shopping on its platform, pointing directly to that tension. Whether agents can cross those boundaries will help determine how much consumer inertia they can actually remove.

If you found even one useful idea in this week’s Intelligence Snacks, we have a small favour to ask. Please forward this newsletter to someone who you think would get something from it too.

Sharing it is one of the simplest ways to support Intelligence Snacks and helps these ideas reach people we might never reach ourselves.

You can find more Snacks from this week’s podcast at intelligencesnacks.com.

See you next week!

Cheers,
Pete & Andy