The case for more rope (and context)
Building agentic systems to tackle larger scope, with more autonomy
Authored by Nico Rakover, Lead Research Engineer
Solver is a cloud-based software engineering agent, optimized for autonomous, asynchronous work. We’ve been building software agents since before “agentic” was in vogue and AI-assisted meant code-completion in the editor. It’s exciting to see more of the industry join us in building increasingly autonomous systems that free up engineers to focus on higher-multiplier activities.
But we’ve watched a dichotomy emerge between autonomy and scope, where products trade off between the complexity of work delegated to AI and how tightly the user is involved. IDE-native systems like Cursor are wading into the agentic space, but largely expect the user to supervise the AI at the granularity of individual edits and commands. And if you close your laptop, Cursor stops working for you. On the other hand, cloud-based agents like Devin and V0 either only support narrow slices of the software stack or ask their users to only throw them short or narrowly-scoped tasks.
A false dichotomy
We believe this tradeoff between scope and autonomy is a self-imposed, artificial tension, and it counterintuitively poses a barrier to the efficacy and adoption of agentic systems.
Clearly, agentic systems that only do small tasks or are marketed as “junior devs” are less compelling than those that can handle complex, multi-day projects. Beyond the lower ROI, it turns out that delegating short asynchronous tasks can have a negative impact on overall productivity – a phenomenon that we’ll explore further in a subsequent post – which in turn applies non-linear resistance on adoption of agentic software systems.
Moreover, an artifact of how agentic systems are built today is that the smaller the delegated task, the less context the system has to help ground and inform its decisions. For this reason, we observe that well-designed software agents can be more reliable on larger tasks.
As a brief detour, consider how we think about scope and autonomy when it comes to human engineers. With experience and tenure, engineers grow to take on more scope with more autonomy. That’s fundamentally how they create greater value and higher velocity for their organizations. Of course, sometimes things don’t go as planned: features flop, systems break or decisions turn out to be suboptimal. But how do most managers and organizations respond? Typically, the realization is that the engineer needed more context, not a narrower view of the world or task.
More scope, more better
Inspired by this observation, we’re building Solver to take on dramatically larger scope and complexity than any other system like it. Not because it nails every small task, but because it misses some. We’re optimizing Solver so that (1) it serves as a thought-partner in thinking through the requirements and complexities of much larger projects and tasks, and (2) it then leverages this grounding context to make better decisions when working through the nitty-gritty details of implementation.
Our Projects feature, currently in preview, is the first incarnation of this philosophy. It provides a two-stage UX:
- Helping users rapidly author structured technical plans for arbitrarily large projects.
- Executing against the plan, asynchronously and autonomously, requiring minimal intervention only when it decides that it can’t confidently self-validate intermediate milestones.
Projects is designed to bootstrap a dense foundation of context via a relatively short but synchronous dialogue with the user. Given a seed of project requirements, Solver iteratively builds up a structured tech plan and, critically, surfaces areas of ambiguity as open questions for the user to answer. This dialogue allows rapidly scoping work that could easily involve multiple human-dev-days of execution. Once the plan is clear, users can one-click execute the entire graph of scoped tasks, such that every minute decision Solver makes during implementation is grounded on its understanding of the full plan.
We’re super excited about our community’s reception of this early iteration of Projects, and for continuing to build our vision for software agents. Looking forward to your thoughts and feedback!