Continual Learning: How AI Agents Get Better With Every Use | Arjun Karanam, Trajectory
- Source
- youtube.com
- Author
- Sequoia Capital
- Date
Context
Arjun Karanam, co-founder of Trajectory, a startup building tools for continual learning (agents that improve with use), says today's agents feel like a new hire on a first day because the experience they generate is thrown away. In this Sequoia Capital talk Karanam sketches a pipeline: interactions, define what the agent should do, then use the records to improve both the model, through reinforcement learning on long traces, and the , the surrounding code and . A fact such as a company being delisted belongs in the harness's context, not trained into the model.
Karanam's advice is to record more. Many companies trace only the main action and discard tool calls and sub-agents, which makes learning from a whole run harder. Thumbs up and down are noisy, Karanam says, because coding-agent users tend to accept everything until a bug surfaces commits later, so edits, undos and retries are the stronger signal. These are a founder's recommendations and the product shown is in beta; the talk reports no results of agents improving this way.
- Arjun Karanam says thumbs-up/down feedback is too noisy, since coding-agent users tend to accept everything. The useful signal is corrective behavior: edits, undos and retries, which products should both prompt for and capture.
- Karanam warns that tool calls returning only 'done' confuse agents and leave nothing to learn from. He recommends tool responses that say what was actually read or written.
- Trajectory sorts feedback by where it belongs. A fact, like a company being delisted, goes to the harness's context and not the weights. A that keeps failing is likely true for everyone, so it goes into the model. One user's preferences stay in context.
- Karanam's eval advice: build evals from real traffic, make every task replayable ('roll outable'), which he calls a big infrastructure challenge, and grade through the real production harness rather than a variant of it.
- On privacy, Karanam says you can avoid training on customer data directly. Instead, sample its distributions, generate synthetic data, and check that the synthetic data matches the real distribution.
- Record — One unit of data in a file or table, such as one supplier's quote; distinct from a line of text.
- Harness — The software that sends a model its context, recognizes tool calls, runs or refuses them, returns results and keeps the record.
- Context and context window — Context is the information available for a generation; the context window is the capacity available to hold it.
- Tool call — Text the model generates in an agreed format that names a tool and gives its inputs.
videoBringing Continual Learning to Every Company: Trajectory’s Ronak Malde
videoCrusoe CEO Chase Lochmiller: Turning Stranded Energy Into Intelligence
videoDavid Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience
videoCracking the Code on Offensive Security With AI ft XBOW CEO and GitHub Copilot Creator Oege de Moor
videoBringing Continual Learning to Every Company: Trajectory’s Ronak MaldeSequoia Capital
videoBuilding the Automated AGI Lab: Core Automation's Jerry Tworek and Rohan AnilSequoia Capital
videoThe Breakthroughs Needed for AGI Have Already Been Made: OpenAI Former Research Head Bob McGrewSequoia Capital
Checking sign-in…
Loading comments…