Interview

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

Source attribution: Original source

Interview details

Publication date: Aug 16, 2026
Duration: 22m

Speakers

Arjun KaranamCo-founder, Trajectory

Viewpoints in this interview

AI progress needs experience alongside model intelligence

Karanam argues that models are rapidly becoming more intelligent, but they still behave as if every interaction were their first day on the job because they do not accumulate experience from prior work.

Arjun Karanam AI Experience Gap Source at 0:39

Agent interactions should become signals for future improvement

Karanam says the work generated through real agent interactions is commonly discarded, but it should instead serve as learning signal that allows agents to improve with continued use.

Arjun Karanam Continual Learning Source at 1:09

Continual learning can create systems that compound with use

Karanam describes continual learning as a path toward faster, better, and cheaper models, with the larger goal of creating systems whose capabilities compound as they are used.

Arjun Karanam Continual Learning Source at 1:32

Feedback should update either models or their surrounding harness

Karanam says continual-learning systems need to determine whether feedback belongs in model weights or in the surrounding harness, depending on the nature of the information being learned.

Arjun Karanam Model Training Source at 2:14

Full agent traces should include tools and sub-agents

Karanam argues that companies should trace the complete execution tree, including tool calls and sub-agents, because omitting those steps makes it difficult to learn from the full interaction.

Arjun Karanam Agent Traceability Source at 3:20

Corrective behavior provides richer feedback than simple ratings

Karanam considers thumbs-up and thumbs-down feedback noisy and argues that edits, undos, and retries reveal more useful information about how an agent should improve.

Arjun Karanam Feedback Signals Source at 3:34

Evals should closely reflect real production usage

Karanam says evaluation environments should resemble both production use and training environments, with evals drawn directly from how users actually interact with the product.

Arjun Karanam Production Evals Source at 3:56

Agent tasks should be replayable for evaluation

Karanam argues that companies should build infrastructure capable of replaying user tasks so prior interactions can be rolled out again for evaluation and improvement.

Arjun Karanam Production Evals Source at 4:13

Harnesses should enable orchestration instead of rigid workflows

Karanam says modern agent harnesses should expose useful primitives such as tools and private information while allowing the agent to orchestrate them rather than forcing predefined execution flows.

Arjun Karanam Agent Harnesses Source at 4:30

Agents should access the same capabilities as users

Karanam recommends aligning agent and user interfaces so that actions available through the product UI are also available to the agent through tool calls.

Arjun Karanam Agent Harnesses Source at 4:51

Tool responses should expose informative execution details

Karanam argues that tool responses should provide meaningful details about what was read, written, or changed rather than returning vague success messages that offer little learning signal.

Arjun Karanam Agent Harnesses Source at 5:02

Open weights enable companies to continually improve owned models

Karanam recommends becoming comfortable operating open-weight models because owning model weights makes continued model improvement possible, while acknowledging operational concerns such as security and safety.

Arjun Karanam Open-Weight Models Source at 5:28

Model routers can match intelligence to task requirements

Karanam expects model routers to play an important role in directing different tasks to intelligence suited to the required capability and recommends experimenting with them.

Arjun Karanam Model Routing Source at 5:43

Continual learning should optimize the whole intelligence system

Karanam defines product intelligence as a system containing models, harnesses, context, and other components, and says continual learning should determine which component should be updated for each new piece of information.

Arjun Karanam Continual Learning Source at 7:37

Learning can avoid direct training on customer data

Karanam says continual-learning systems can derive distributions from customer data, generate synthetic data from those distributions, and evaluate distribution alignment without directly training on the underlying customer data.

Arjun Karanam Customer Data Privacy Source at 8:27

Corrective feedback gives stronger reward signals than dissatisfaction

Karanam distinguishes feedback that only indicates failure from feedback that demonstrates a successful correction, saying the latter provides much stronger evidence about the appropriate reward and desired output.

Arjun Karanam Feedback Signals Source at 9:11

Feedback scope determines where learned information should live

Karanam says globally relevant behavior can be trained into the model, while individual user preferences should remain in context, with some information potentially applying at organization or customer scope.

Arjun Karanam Learning Scope Source at 9:35

Continual learning is most compelling near capability frontiers

Karanam says he is especially interested in tasks near the boundary of current model capabilities, where unsuccessful user requests can become training signals that expand what the system can subsequently accomplish.

Arjun Karanam Frontier Tasks Source at 10:31