Exploring Proactive Agents at AirJelly: Working the Way You Do

A conversation ends. The work often continues.

Plans still need to move forward. An email may bring a change, a meeting may be rescheduled, and a question discussed moments ago may find its answer in another application. For an agent that aims to participate in this work, messages from the user are just one way to understand what is happening.

In January 2026, OpenClaw went viral.

One important lesson it offered was that an agent can keep participating in work even when the user has not asked a new question. Through scheduled tasks and heartbeats, an agent can check on relevant matters and decide whether to help. This made the idea of an assistant that remains present more tangible. It also opened up a direction worth exploring: when events beyond the passage of time can trigger an agent, how might it participate in our everyday work?

At AirJelly, we envision proactive agents that gain new context as work unfolds, use the user's goals and habits to judge when to act, and draw on the right capabilities to help.

This article explores one possible architecture and the ideas behind it.

Proactivity starts with the user

It is easy to describe proactivity as a list of features: meeting reminders, email organization, task follow-ups, and suggestions for what to do next. Yet the same feature can have very different value in different people's work.

Some people want to hear about important changes immediately; others prefer to handle messages in batches. Some like to review a draft before discussing it, while others want to agree on a direction before any work begins. Even the same person may welcome different levels of involvement while exploring a problem, concentrating on creative work, or delivering a result.

These differences reflect how people organize their work: how they allocate attention, make decisions, collaborate with others, and decide what to delegate to an agent.

A proactive agent therefore needs to consider the user's goals, preferences, and present circumstances when deciding how to participate. An email about an upcoming review might be a reason to prepare materials, update an arrangement, or wait until the user finishes what they are doing.

We want that judgment to remain flexible. Users should be able to express what they want an agent to pay attention to, what it should take on, and when it should stay quiet, without having to adapt to a predefined workflow.

Changes in the environment become opportunities to collaborate

Scheduled tasks make time a trigger. Extending that idea, many changes in the environment can give an agent a reason to reconsider the current situation.

An approaching meeting, a new email, a desktop notification, a change in a connected device's state, or an action completed in an application may all be relevant to ongoing work.

Desktop activity is particularly interesting. As people read, discuss ideas, write code, and adjust plans, the context of their work spreads across applications. If an agent can understand that activity within the boundaries the user permits, it may reduce the burden of repeatedly explaining the background or assembling information again.

An event only tells us that something may have changed. Whether it matters, which goal it relates to, and whether action is needed all depend on context. A trigger creates an opportunity for judgment; the user's work gives that opportunity meaning.

We want an agent's attention to follow those relationships: watching relevant arrangements while a review is being prepared, respecting concentration during writing, and changing its involvement when a task is complete.

An agent that remains involved

These ideas lead us to an architecture composed of triggers, a main agent, tools, and sub-agents. User context informs decisions, while the results of actions and feedback from the user shape subsequent judgments.

The main agent carries forward the context of the work, judges whether a change deserves attention, and decides what should happen next. It can gather more information, ask the user a question, invoke a tool, or delegate work that needs further exploration to a sub-agent.

Sub-agents work toward specific objectives while the main agent continues to attend to the user and changes in the environment. Anthropic's multi-agent research system also uses this orchestrator–worker division. For proactive agents, it offers an important design perspective: while a task is being carried out, someone still needs to attend to whether its goal has changed. Anthropic's multi-agent architecture

For example, if a review is rescheduled again, the preparation for it should change accordingly. New information needs to affect work in progress, rather than merely appear in the next reply. Ongoing collaboration between the main agent and its workers makes changing direction, adding context, or stopping a task part of that relationship.

Here, continuity means that collaboration can extend beyond a single conversation. After a response ends, the user's goals, unfinished work, and subsequent changes remain connected. The agent's involvement should carry those connections forward.

Every plugin is a two-way connection

Adapting to different tools and working habits requires capabilities that are composable and pluggable.

In the architecture we envision, every plugin adds capabilities in two directions: a plugin can trigger the agent through events, and the agent can invoke the tools that the plugin provides.

A calendar plugin can trigger the agent when a meeting is approaching or an arrangement changes. It can also be called to query, create, and update calendar events. An email plugin can provide new input when a message arrives, as well as tools for reading and replying. A device plugin can report state changes and provide corresponding controls.

Every plugin can therefore be both the starting point for collaboration and a participant in the actions that follow.

More importantly, these capabilities can work across plugins. A change communicated by email might prompt the agent to check the calendar and then delegate the preparation of relevant materials to a sub-agent. The plugin that introduces an event does not determine which plugin the agent uses next. The user's goals and current context guide that choice.

As plugins are added, the environment the agent can understand and the actions it can take expand together. Users can choose capabilities around their own work without being confined to a particular application, a category of tasks, or a fixed workflow.

Let the user set the pace

The value of a proactive agent ultimately depends on whether it provides appropriate help. Sometimes that means preparing materials in advance or following up on unfinished work. Sometimes it means understanding the situation and choosing not to interrupt.

This kind of collaboration needs to leave room for the user to express preferences and make adjustments. Users should be able to change what the agent pays attention to, correct its understanding of the work, and decide which actions should continue and which should stop.

A composable architecture provides a foundation for that flexibility. Triggers bring changes to the agent's attention, user context informs judgment, and tools and sub-agents carry out actions. These parts connect around the work the user is trying to move forward.

This is one vision for proactive agents at AirJelly: agents that participate more naturally in ongoing work, help in ways that follow the user's goals and working habits, and adjust their involvement as circumstances change.