Whether you're a scientist brainstorming research ideas or a CEO trying to automate a human resources or finance task, you'll find that artificial intelligence tools turn out to be the assistants you didn't know you needed. Many professionals particularly are tap into the skills of semi-autonomous software systems, so-called AI agents, which might depend on AI at certain points to resolve problems and complete tasks.
AI agents are particularly effective when using large language models (LLMs) because these systems are powerful, efficient, and adaptable. One option to program such technology is to explain in code what you wish your system to do (the “workflow”), including using an LLM. If you were a software company trying to refactor your legacy codebase to make use of a more modern programming language for higher optimizations and security, you possibly can create a system that uses an LLM to translate the codebase file by file, testing each file as you go.
But what happens when LLMs make mistakes? You want the agent to back down and check out again, incorporating lessons from previous mistakes. Coding can require as much effort as implementing the unique agent. If your system for translating a code base contained 1000’s of lines of code, you’d make 1000’s of lines of code changes or additions to support the logic for backtracking when LLMs make mistakes.
To save programmers effort and time, researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) and Asari AI did this developed a framework called “EnCompass”.
With EnCompass you not must make these changes yourself. Instead, EnCompass robotically backtracks when running your program if LLMs make mistakes. EnCompass may create clones of this system runtime to run multiple experiments in parallel in quest of the very best solution. In general, EnCompass scans the various possible paths your agent might take given the various possible outputs of all LLM calls and appears for the trail where the LLM finds the very best solution.
Then all it’s essential do is annotate the locations where it is advisable to retrace or clone this system runtime and record any information that could be useful for the strategy used to go looking the various possible execution paths of your agent (the search strategy). You can then set the search strategy individually – you’ll be able to either use one which EnCompass provides by default, or implement your individual custom search strategy if mandatory.
“With EnCompass, we separated the search strategy from the underlying workflow of an AI agent,” says lead writer Zhening Li '25, MEng '25, a doctoral candidate in electrical engineering and computer science (EECS) at MIT, a CSAIL researcher, and a research advisor at Asari AI. “Our framework allows programmers to simply experiment with different search strategies to search out the one that offers the AI agent the very best performance.”
EnCompass has been used for agents implemented as Python programs calling LLMs and showed noticeable code savings there. EnCompass reduced the coding effort required to implement search across agents by as much as 80 percent, akin to an agent to translate code repositories and discover digital grid transformation rules. In the long run, EnCompass could enable agents to handle large-scale tasks, including managing massive code libraries, designing and conducting scientific experiments, and creating blueprints for rockets and other hardware.
branch
When you program your agent, you highlight specific processes – e.g. B. Calls to an LLM – where results may vary. These annotations are called “branch points.” If you imagine your agent program generating a single story arc, adding branch points turns the story right into a choose-your-own-adventure story game, where branch points are places where the plot branches into multiple future storylines.
You can then set the strategy that EnCompass will use to navigate the story game to search out the very best possible ending to the story. This can include starting parallel threads of execution or backtracking to a previous branch point when you're stuck in a dead end.
Users may plug-and-play some common search strategies that EnCompass provides out of the box or define their very own custom strategy. For example, you possibly can select Monte Carlo tree search, which builds a search tree by balancing exploration and exploitation, or beam search, which keeps the very best results at each step. With EnCompass, you’ll be able to easily experiment with different approaches to search out the very best strategy and maximize the likelihood of completing your task successfully.
The coding efficiency of EnCompass
So how code efficient is EnCompass when adding search capabilities to agent programs? According to the researchers' findings, the framework dramatically reduced the quantity of effort that programmers had so as to add to go looking of their agent programs and helped them experiment with different strategies to search out the one which performed best.
For example, the researchers applied EnCompass to an agent that translates a code repository from the Java programming language, commonly used for programming apps and enterprise software, into Python. They found that implementing search with EnCompass—mainly adding branch point annotations and annotations that record how well each step worked—required 348 fewer lines of code (about 82 percent) than implementing it manually. They also demonstrated how EnCompass allowed them to simply try different search strategies, with a two-stage beam search algorithm being the very best strategy and achieving a 15 to 40 percent increase in accuracy across five different repositories with a search budget equal to 16 times the LLM calls made by the non-search agent.
“As LLMs turn out to be more integral to on a regular basis software, it would turn out to be increasingly vital to know find out how to efficiently develop software that leverages its strengths and works around its limitations,” says co-author Armando Solar-Lezama, MIT Professor of EECS and CSAIL principal investigator. “EnCompass is a very important step on this direction.”
The researchers add that EnCompass targets agents where a program determines the steps of the high-level workflow; The current version of their framework is less applicable to agents which are fully controlled by an LLM. “With these agents, there isn’t a program that specifies the steps after which uses an LLM to execute those steps, but slightly the LLM itself decides every thing,” says Li. “There isn’t any underlying programmatic workflow, so you’ll be able to run inference time search on anything the LLM invents on the fly. In this case, there’s less need for a tool like EnCompass that manipulates the execution of a program through search and backtracking.”
Li and his colleagues plan to increase EnCompass to more general search frameworks for AI agents. They plan to check their system on more complex tasks with a purpose to further develop it for real use, including in corporations. Additionally, they assess how well EnCompass helps agents collaborate with humans on tasks akin to brainstorming hardware designs or translating much larger code libraries. Currently, EnCompass is a robust constructing block that permits people to more easily tinker with AI agents to enhance their performance.
“EnCompass comes at the proper time as AI-driven agents and search-based techniques begin to reshape software development workflows,” says Professor Yiming Yang of Carnegie Mellon University, who was not involved within the research. “By neatly separating an agent's programming logic from its inference-time search strategy, the framework provides a principled option to examine how structured search can improve code generation, translation, and evaluation. This abstraction provides a solid foundation for more systematic and reliable search-driven approaches to software development.”
Li and Solar-Lezama co-wrote the paper with two Asari AI researchers: Caltech professor Yisong Yue, an advisor to the corporate; and lead writer Stephan Zheng, who’s founder and CEO. Their work was supported by Asari AI.
The team's work was presented in December on the Conference on Neural Information Processing Systems (NeurIPS).
