Aug 13, 2026, 9:00 AM ETConference Presentation, Keynote

Continual Learning: How AI Agents Get Better With Every Use | Arjun Karanam, Trajectory

Mission: Trajectory, founded by Arjun and Ronak (co-founders), is building a platform to close the "experience gap" in AI by enabling models to learn from real-world interactions over time.
Core Problem: Current models exhibit high intelligence (IQ) but lack accumulated experience, acting like a "Terence Tao on his first day" rather than a seasoned expert.
Market Signal: Approximately 100 trillion tokens are generated daily by AI agents, yet the vast majority of this real-world work and user feedback is discarded.
Product Vision: A system where agents compound in capability with use, resulting in faster, cheaper, and more effective models that improve continuously based on user interactions.

Wish 1: Traceability
- Capture the entire decision tree, including sub-agent calls and tool usage, rather than just the main action.
- Elicit and capture "corrective behavior" (edits, retries, undos) rather than relying solely on noisy binary feedback (thumbs up/down).
Wish 2: Evals
- Align evaluation environments directly with production traffic and user requests, including frontier tasks users attempt.
- Ensure every task is "rolloutable," allowing for the replay of user actions to verify agent performance.
- Grade agents using the exact, real harnesses deployed in production, avoiding simplified or simulated environments.
Wish 3: Harness Optimization
- Shift from enforcing rigid execution flows to providing flexible primitives (tools, private data) for agents to orchestrate.
- Make the agent interface a 1:1 mapping of the user interface, enabling agents to execute via tool calls exactly what users do visually.
- Ensure tool responses are informative (e.g., confirming specific DB writes or reads) to provide dense learning signals for the agent.
Wish 4: Model Strategy
- Transition to open-weight models to enable ownership of weights and continuous fine-tuning, despite security and provisioning complexities.
- Implement model routers to direct intelligence to the specific model best suited for a given task's capability level.

Model Spec Definition: Research focuses on extracting user interactions to generate precise "model specs" that define desired agent behaviors.
Dual-Learning Surfaces:
- Models: Utilizes Reinforcement Learning (RL) algorithms like SDPO to learn from long interaction traces.
- Harnesses: Routes specific knowledge (e.g., "Company X is delisted") to context or harnesses rather than retraining static model weights.
Signal Hierarchy:
- High Confidence: Direct corrections (edits, retries) provide clear reward signals for training.
- Low Confidence: Negative feedback (e.g., session drop-off, generic complaints) is used to penalize behavior without assuming a specific correct alternative.
Scalability of Updates: Decisions on what to update (model vs. harness vs. context) are treated as a scientific optimization problem, abstracting away manual configuration for the user.

Differential Privacy: The approach avoids direct training on customer data by sampling distributions and synthetically generating training data to protect customer-specific arrangements.
Per-Organization Learning: Feedback hierarchy is established based on pertinence: globally accurate errors (e.g., tool failures) train the model, while user-specific preferences (e.g., "never use Subagent Y") are handled via context or per-org customization.

Post-Training Automation: The platform automates ~50 complex knobs typically required for post-training, reducing model training and validation to approximately 15 minutes of active user work.
Interface: Features an agent-driven interface that simplifies importing training data, evaluating models against benchmarks (e.g., "Harvey"), comparing versions, and deploying.
Target Audience: Designed to empower companies to build internal expertise in model ownership rather than outsourcing intelligence to consultants.
Frontier Focus: The system is most effective on tasks at the edge of current capabilities, allowing models to learn specific behaviors that users previously could not achieve.