What We ACTUALLY Build at Teammates

We often get asked: “So what are you actually building at Teammates?” It’s sort of a strange question – I doubt the engineers who build Block Blast! get asked that question. It’s pretty clear they’re writing software code of some sort to build an iPhone game. Most people can roughly understand that. Even if they aren’t super technical. 

So why do people look so puzzled when they ask what we’re building?

It’s surprising – folks are genuinely curious, but aren’t even sure how to ask the question. We get a mix of “Are you writing code?”, “Are you writing prompts?”, “Are you like using ChatGPT?”, “Do you have servers??”

And to be honest, I haven’t been sure how to answer their question. I mean, of course we write code, and prompts, and use frontier models, and run servers. But that doesn’t explain where all our time and energy is going, and it also doesn’t explain why I keep saying that Teammates is, by far, the hardest thing I’ve ever built.

Which is why I was excited to read Harrison Chase’s recent article entitled The rise of "context engineering". Context Engineering a new term, in which he pretty clearly articulates a new discipline of engineering that is emerging. To quote the article:

Context engineering is building dynamic systems to provide the right information and tools in the right format such that the LLM can plausibly accomplish the task.

Early on, developers focused on phrasing prompts cleverly to coax better answers. But as applications grow more complex, it’s becoming clear that providing complete and structured context to the AI is far more important than any magic wording.

It’s the best explanation I’ve read yet about where the industry is heading, and what the Teammates engineering team does all day. 

To be reductive, a Teammate is a container for context: a Teammate has a job description, personality, short and long term memory, all its past work assignments, feedback, conversation history from slack, runbooks, company knowledge base, skills, and software tools. And our job as product engineers is to take this massive amount of dynamic, ever-changing, personalized context and turn it into actionable prompts that can drive agentic workflows. 

And when we do our jobs well, we start to see the characteristics of Teammates start to emerge: autonomy, intuition, collaboration, and evolution. These traits arise from context engineering; given all the interactions and history and feedback and conversations that they’ve even had, when combined in ingenious ways and fed into an LLM (or more accurately, our agentic framework), produce software like you’ve never experienced.

I would also posit that this is why code generation is one of the early smash hits of agents. In the early days (read: 6 months ago), folks would say “coding is a good LLM task because the vocabulary is finite and code can be tested for correctness.” Which is certainly true, but that doesn’t consider context. And a software project is a highly structured set of data that can create a very clear context. The codebase, the files, the terminal output, the programming language best practices, the output of linters, etc. All of these things feed into the context, and notably, it’s a constrained problem space. The difference between a vibe coder and a principal software engineer wielding these tools is their ability to manipulate and provide better and better input context: the right files, the right prompt, the right terminal access, etc.

But back to Teammates… when a customer designs a Virtual Recruiter and asks for help reviewing resumes, it’s the context that will take her from a generic ChatGPT output to understanding the hiring manager’s priorities and personal preferences (“I hate job hoppers”), the nonnegotiable aspects of the role, whether we’ve seen this candidate 4 times already, how to advance stages in the applicant tracking system, how the recruiter likes to format their candidate feedback, etc. So when a new candidate applies, they can get right to work with all the context and knowledge they’ve built up over time.

And when a Virtual Operations Manager logs into an admin portal to fill out insurance details, or a Virtual Customer Success Manager searches for the latest news articles that could impact their book of business, or a Virtual Sales Specialist helps book a demo call over email, they all have context. The context they need to not just be a simple prompt to coax a good LLM output. But a series of dozens or hundreds of prompts, all dynamically generated based on hundreds or thousands of past interactions, conversations, constructive feedback, history, mistakes, compliments, memories, knowledge, and more.