Skip to content
AI.info

The Pulse

LangChain CEO calls for agents to pursue goals, not just tasks

LangChain CEO Harrison Chase says AI agents are moving from discrete assignments to ongoing responsibilities. He outlines requirements for choosing work, seeking feedback and deciding when to act.

LangChain CEO calls for agents to pursue goals, not just tasks

AI agents are moving from completing discrete tasks to pursuing ongoing goals, according to LangChain CEO and co-founder Harrison Chase. In a post on November 16, he outlined what that shift would require: agents that choose their own work, seek clarification, measure progress and decide when to act.

Instead of asking an agent to draft a blog post or fix a bug, a user might give it responsibility for maintaining a blog or helping make a product successful. Chase describes the change as moving from a specified action to an outcome, and from a bounded assignment to an ongoing responsibility.

Choosing the work

In that model, the user sets the goal and the agent breaks it into tasks, carries out some of them and updates its list as it goes. That gives the agent a role in deciding what work to do, rather than simply executing a task the user has selected.

Chase says agents need to generate varied possibilities rather than repeatedly proposing the same tasks. He also calls for a way for users to see what work an agent is doing, such as a to-do list, alongside prompting that encourages varied ideas.

The tools an agent needs can also be harder to predict when it is pursuing a broad objective. Sending an email points to an email tool and perhaps a calendar tool; reducing a person's coordination burden could involve email, text, search, a browser or other tools at different times.

Chase calls for controls over which tools an agent can use. Background agents also need a way to ask people for clarification, he says, whether through Slack or a dedicated interface, rather than relying on an active chat session.

Feedback and progress

An agent working toward a goal needs a way to assess how it is doing. Chase suggests that agents could check a metric for improvement or receive explicit feedback from a person, with both approaches available together.

Metrics can be easier for an agent to track automatically, but he warns against blindly optimizing them. Goals should be designed so that feedback is easy to provide and users can correct the agent's interpretation of success, he says.

The post also sets out a role for memory across separate runs. When a user gives an agent an individual task, the user can supply the relevant context each time; for an ongoing responsibility, Chase says the agent needs to carry decisions and feedback forward.

For example, rejecting a blog topic should influence future topic selection rather than only the current draft. That requires retaining and using relevant decisions across runs.

When to act

Agents assigned ongoing goals need not run continuously, but they do need a way to resume work. Possible triggers include periodic heartbeats, recurring schedules, one-off reminders and event streams.

Those triggers also need stopping or pausing rules. An agent that asks the same question at every thirty-minute heartbeat would offer a poor experience, Chase says; sometimes the appropriate action is to do nothing.

In his account, delegating a goal means giving the agent discretion over what to do and when, while the user sets direction and boundaries and reviews progress. He presents those requirements as practical considerations for building and interacting with goal-directed agents.

Source

Companies in this story

Related stories

Explore

More articles