The Great Unbundling of Intelligence
Why I think specialized agents will beat general-purpose ones, and what actually gets unbundled
At Nori, we started by running support tickets through general-purpose coding agents like Codex and Claude Code. It worked reasonably well, especially at first. The agents could inspect an issue, search through the codebase, and often come back with something useful. The problem was that none of this compounded. Months later, we were still getting roughly the same results. Ticket five hundred was handled with the same generic context as ticket one.
This led us to build a specialized support agent in-house. It triages tickets automatically through webhooks, lives natively in Slack, debugs issues as they come in, and can respond directly to users. Everything in its context serves one job: managing our support queue. Because the task is narrow, we can evaluate its work, capture where it went wrong, and use that feedback to improve the next result.
That experience changed how I think about AI products. General-purpose agents are a great place to start because they let you test an idea without building much infrastructure. But once you understand the work, there is a real advantage in building a system around that specific job. The general agent helped us prove the workflow. The specialized agent gave us something we could continue improving.
This is happening elsewhere
Two recent product launches made me think our support experience was part of a larger change.
Harvey released Tenet, a model post-trained for legal reasoning. Harvey did not train it from scratch. It started with Kimi K3, an open-weight model, then trained it with synthetic examples, public legal material, and work from legal experts. On Harvey’s legal benchmarks, Tenet beats the general-purpose models it replaces inside the product while costing less to run.
Harvey also plans to let each law firm train the model further on its own work. If that works, two firms using the same product will end up with models that behave differently. Each model will reflect how its firm writes, reviews, and makes decisions.
Around the same time, Wispr Flow announced Canto, a speech model trained for real-world dictation. It focused on the noisy rooms, accents, and interruptions that clean recordings leave out. Wispr says that in the hardest conditions, its word error rate falls from more than 30% to between 5% and 10%.
Legal reasoning and speech recognition are very different problems. The approach is similar: start with what general models can do, then teach a focused system about the work people actually need it to perform.
The work is more specific than the model
A general model can know an extraordinary amount about contracts. It still does not know how a specific firm drafts a purchase agreement, which clauses its clients care about, or what its partners consider finished work.
This is true outside law. A support team has its own definition of a good response. A medical practice has its own standards for a useful chart. An engineering team has opinions about when to repair a system and when to replace it. Much of this knowledge never appears in public training data. It lives in past decisions, corrections, and conversations between people who do the work.
A focused agent has practical advantages too. It needs fewer tools and less context. There are fewer ways for it to wander into the wrong part of a system. You can also evaluate it against a real standard. “Did this agent find the cause of the ticket and propose the correct response?” is a question we can answer. “Was this generally helpful?” is much harder to improve against.
Once the standard is clear, normal use becomes useful feedback. You can save the result, the correction, and the final outcome. Over time, those records become something the general model provider cannot build for you because it does not see your work.
What gets unbundled
SaaS split large software suites into separate tools for sales, finance, support, and every other department. I think AI will do something similar to the models inside those tools.
For the last few years, companies have bought intelligence in one bundle: one model and one API for many unrelated tasks. I expect more products to separate that bundle. They will use a strong base model, teach it about one kind of work, place it inside a workflow, and measure it against the people who already do that work well.
Foundation model companies will continue to push the general frontier, and that work remains important. But access to a frontier model will not make one legal product different from another. The difference will come from the legal work each company can learn from, the standards it can measure, and the workflow it builds around the model.
In that sense, the model itself may become less important than what a company can teach it about its work.
Where the argument breaks
A specialized agent can become a dead end. It can do familiar work very well and then fail when a ticket crosses a boundary nobody expected. A narrow model may also be a poor investment when the task changes often or when there are too few examples to learn from.
That is why our support agent should not be a wall. The narrow system can handle the work it knows and pass unusual cases to a frontier model or a person. The general model remains valuable because it can reason across boundaries. The specialized agent is valuable because it understands the repeated work inside them.
I do not know how far this goes. A future general model may need much less help to learn a company’s context. Better long-context systems may remove some of the advantage of smaller prompts. Some work may never repeat enough to justify a dedicated agent.
But I have become convinced of one thing: “general-purpose” describes the starting model, not the finished product.
The useful question is not whether a general model can perform a task. It often can. The question is whether your system will be better after it has performed that task five hundred times.
Thanks for reading🙂 If you have thoughts about where this argument holds or breaks, e-mail me at [email protected].