Some Thoughts On Agents -
Introduction
When I first started writing code, I built agents that felt dangerous and somewhat illicit. I was doing things that would have been termed malicious: creating bot networks capable of large-scale automation. ”Botting” as it’s called. Those bots navigated the internet, scraped data, executed tasks, and evaded authentication.
Of course, agents have become more mainstream since then –
The big change is that bots are now easier to create. APIs provide easy access and connectivity, along with improvements in hosting, data loading, and memory management.
But way more profound was the release of OpenAI’s Codex and other tools that turn language into code.
It’s opened up people to create and train agents to do all sorts of things — and unlocked a future where agents can reason and learn. It’s caused a massive explosion in the number of agents and agent companies, leading to a diversification and also specialization within the industry.
I think recently we’ve all felt a bit of agent fatigue. “Agent” has become a catch-all term for software that executes a specific task or workflow. And since most software is created to execute a task or workflow, the word is everywhere.
However, we’re still in early innings of agent capabilities, and we see huge opportunities within the space.
Natalie and I at Chapter One wanted to share some high-level thoughts on what excites us about agents and where we think the investment opportunities are.
What excites us about agents
Agent Experience
For the past 10 years, big tech has been focused on user experience. The conversation in every product meeting is: customer needs, habits, and behaviors. Most of the Internet and web interfaces were designed with humans in mind.
But now – the agents are the users. It creates a new problem space: “agent experience”.
Researchers at Princeton recently came out with a paper titled: SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
This paper shows how language models can perform better on a custom agent-computer interface improving an agent’s work in coding, navigating repositories and running tests.
Inspired by human-computer interaction (HCI) studies on the efficacy of user interfaces for humans, we investigate whether LM agents could similarly benefit from better-designed interfaces for performing software engineering tasks
Right now, agents are best at backend engineering, developer operations, developer setup, database setup, running Kubernetes, and data analysis.
We’re interested in infrastructure that’s built specifically for agents and agent experience. To name a few:
Browserbase: Browsers built for AI agents.
Browserless: Provides headless browser automation without the need for GUI.
Langchain - Framework that bridges the gap between LLM capabilities and automated action.
Apify: A platform where developers build, deploy, and publish web scraping, data extraction, and web automation tools.
Anon: Empowers AI developers to build user-permissioned integrations for sites without APIs.
NPi AI: Seamlessly integrate GPTs, AI agents and assistants.
Lutra - automates manual data tasks across apps and the Internet, transforming any data source into structured formats with no engineering.
We believe there is the most white space here for investment opportunities.
Communication Automation
You can really get into a rabbit hole imagining the future of communication amongst virtual agents. It gets abstract pretty quickly. It is discussed a decent amount, but let’s quickly dig in.
A world where: a whole stable of agents can negotiate, look through contracts, and give a first handshake when talking to a booking agent — even date for you. Bumble’s founder, Whitney Wolfe Herd, said she believes in a world where agents will make a first pass at dating capability.
There are a tremendous number of meaningful interactions between people that will probably be automated. It’s not like we’ll have 1-2 of these agents running for us; we’ll have dozens. It’ll be a tiered system: a layer of agents, a layer of quasi- and semi-automated with humans in the loop, and then the human layer.
Lower, less significant interactions become more automated and save effort for more significant types of communications and actions.
This vision seems the farthest away from product-market-fit – and we believe companies with a consumer go-to-market will have a greater uphill battle than developers and enterprise agent use cases. However, there are a lot of companies that are building interesting and innovative solutions in the space. Some that excite us:
Zeta Labs - Jace, an AI assistant that can book reservations and start its own company. Although the browsing speed is still slow and not always reliable, Jace can set up its own LLC for a math tutoring company, send emails, and put together problem sheets.
MultiOn - AI agents that plans and books gathers for you.
Clara Labs - virtual employee that schedules your meetings.
Co-Pilots
Varun Shenoy wrote a blog post Where Are the Good AI Products? arguing that “most people have been yet to be impacted by any sort of AI whatsoever.” – with the exception of ChatGPT & Copilot. And I’ve thought about this truism a lot since that post.
Almost overnight, every engineer I know started using a copilot. Engineers are extremely particular, making this fact even more incredible. There are two different camps here: a view where software is created through a chat interface, and the chat turns into a codebase. On the other hand, you have copilots which are plugins that work alongside software engineers.
A huge amount of money has been invested in this area over the past year and a half.
Of course, it’s because the market size is massive and the efficiencies are tangible –
Personally, I can’t wait to get my hands on Cognition’s Devin.
Some companies to call out:
Cognition - Devin is the first fully autonomous AI software engineer.
Factory - Factory helps organizations automate and optimize their software development lifecycle with autonomous, AI-powered systems called Droids.
Cursor - The AI Code Editor.
Codeium - The modern coding superpower.
Tabnine - AI code completion tool that helps streamline the full software development life cycle.
Codium AI - CodiumAI uses AI to generate meaningful tests for code.
V0 by vercel - V0 generates UI from simple text prompts and images.
Pythagora - OSS dev tool that builds entire apps from scratch by talking to you.
Outro
In terms of adoption and opportunities, we believe the biggest opportunities will be in developer tooling and B2B/enterprise.
For developers, we’re looking ahead at a future where the tools nail down agentic loops and in the background, agents can learn context and deeply understand code bases. On the B2B/enterprise side, we have further to go – search, chat, and knowledge management tools are just gaining traction – so we imagine there will be slower adoption of the second-order tasks.
This is all to say that we’re excited about agent innovation since we’ve only scratched the surface of what agents can do, and the work they can do for us.
While I have nostalgia over my botting days and wish all the agent browsers existed 10 years ago, I look forward to seeing the growth of all the new agent software hitting the market.
If you’re building agent infrastructure or want to hear war stories about botting, reach out to us! Chapter One invests $500K to $2M at the pre-seed and seed stages.