Eran Yahav says some complicated businesses report that new engineers take six to nine months to become productive.
“Uh, some of the more complicated businesses will tell you six to nine months before the engineer really becomes, uh, productive.”
Person
Eran Yahav
1 episodes · 19 quoted claims
Eran Yahav says legacy enterprise systems can contain millions to hundreds of millions of lines of code.
“This is, uh, you are entering into this legacy system, potentially consisting of like millions, tens or hundreds of millions of lines of code”
Eran Yahav describes the agent stack as three components: the language model, the agent and the context engine.
“There's some component to that, but the third layer that is really important in enterprise is what you called the context layer, or what we call the context engine.”
Eran Yahav says large language models are trained on open source code and the internet and know nothing about a particular company.
“They don't know anything about my particular business, my particular bank, my particular enterprise, right?”
Eran Yahav says a context engine pre-computes a map of dependencies and relationships between an organization's components.
“it does pre-computation and builds a map of the organization to understand the dependencies and relationships between components.”
Eran Yahav says a context engine uses internal context agents that continually crawl the organization to build its map.
“These context agents, their job is to do the homework, their job is kind of like constantly crawl the organization”
Eran Yahav compares the context engine's pre-built map to a highway map that lets agents skip irrelevant areas.
“it pre-builds, call it the highway map of the organization for the agent.”
Eran Yahav says a task agent without context must search the whole organization to find the blast radius of a vulnerability.
“typically, if a vulnerability lands and you give it to a task agent with no context, it will have to start digging through the org”
Eran Yahav says many Tabnine customers want the context engine to run inside their perimeter, behind a firewall or air-gapped.
“they would like the context engine to run inside their perimeter.”
Eran Yahav recommends starting by giving the context engine access to code, Jira task management and Confluence design documents.
“giving context engine access to code and to, um, um, I would say something like a Jira task management or Confluence design documents”
Eran Yahav says that in some cases richer context raises agent success rates two times.
“In some cases, we see two times on the success rate”
Eran Yahav says some cases show about 80% improvement in efficiency measures such as token consumption and speed.
“In some cases, you you can see 80% improvement, right?”
Eran Yahav gives an example where an organization has fourteen ways to get employee details and an agent may find an outdated one.
“But in reality, there are fourteen different ways, uh, to get the employee details, and agent just bumped into one of them”
Eran Yahav says code is already being generated faster than humans can review it.
“So we're already kind of generating code faster than any human can review it”
Eran Yahav says it is not unusual for people to spend thousands of dollars on agentic workflows that achieve nothing.
“It's not unusual to see people spending thousands of dollars getting nowhere”
Eran Yahav advises architects to give agents swim lanes by modularizing systems so agents don't interfere with one another.
“you need to give agents kind of swim lanes in which they can operate without kind of interfering with one another”
Eran Yahav compares an AI agent to a Ferrari that burns fuel driving in circles without a map.
“the agent itself is like this very powerful car. It's like a Ferrari, it can go really, really fast, but if it doesn't have a map”
Eran Yahav says the context engine gives architects and engineering VPs visibility across projects and silos.
“Context engine also provides you, maybe for the first time, full visibility of what is going on across different projects”
Eran Yahav says Tabnine's context engine currently focuses on software engineering tasks.
“we are focused on software engineering agents, so we are only looking at software engineering tasks at this point.”