SUMM

Imbue is focused on building robust and useful AI agents, with current work emphasizing software agents.
Their main product, Sculptor, is designed to bridge the gap between AI-generated prototype code and production-ready code.
Existing AI coding tools often produce code that needs substantial manual review before shipping, especially for larger codebases.
The presentation covers technical decisions and lessons learned from developing Sculptor.

Many foundational coding agent improvements (context windows, speed, output parsing) will be solved as models advance.
Imbue prioritizes domain-specific features over generic improvements.
The core problem Sculptor addresses is code review: helping users trust AI-generated code by identifying potential issues in diffs.

Sculptor introduces a third option beyond manual review or blind merging by using an AI agent to inspect code changes for problems (e.g., race conditions, exposed secrets).
High-quality code is defined as code with few or no defects, aligning with academic standards.
The goal is early problem detection in the development process rather than relying solely on late-stage (pull request) reviews.
Sculptor provides synchronous, immediate feedback after code generation.

Four main prevention strategies are implemented: learning, planning, writing specs, and enforcing a strict style guide.
Sculptor assists users in researching existing technologies and solutions to minimize duplicating existing work.
Users are encouraged to make coding plans before generating code, with meta-parameters and system prompts guiding agent behavior.
Specs and documentation are prioritized and managed to stay current, with tools detecting outdated or inconsistent docs.
A strict, possibly AI-specific, style guide helps both agents and humans write safer, more maintainable code—e.g., promoting immutability to reduce race conditions.

Even with prevention, some issues will occur; Sculptor integrates three main detection mechanisms: linters, tests, and LLM-based checks.
Automated tools (like Ruff, Pylint, etc.) are leveraged for error detection and AI can fix common linting issues automatically.
Sculptor tracks pre-existing and new issues to avoid introducing new errors, even if the codebase isn't perfectly clean.
Testing is integrated deeply; AI-generated tests make it feasible to maintain thorough coverage.
Key testing practices include:
- Functional, side-effect-free code for safer test automation.
- Both "happy path" and bug-seeking ("unhappy") unit tests.
- Willingness to regenerate or refactor test suites to stay current.
- Emphasis on integration tests for user-facing behavior validation.
- Monitoring test coverage to boost confidence in the correctness of code changes.
- Running tests in sandboxes with minimal secrets to mitigate risk.

LLMs are used to check a wide variety of code attributes, including uncommitted changes, architectural rule violations, missing or untested specs, and more.
Sculptor allows for extensible, custom checks aligned with organizational best practices.

Most of the challenge lies in identifying, rather than fixing, problems; once well-defined, issues are typically easy for AI agents to address.
Multiple attempts with sandboxed agents can efficiently produce acceptable solutions.
Safe and isolated execution environments enable scalable, parallel problem-solving.

The landscape for AI-driven developer tools is rapidly evolving; upcoming advancements include automated debugging, QA, profiling, visual-to-code tools, and enhanced context-aware tools.
Imbue is open to integrating adjacent developer tools with Sculptor to improve the comprehensive developer experience.
As integration across tools improves, development processes are expected to become much easier and more accessible.

Sculptor is available for sign-up at imbue.com.
Imbue is hiring and eager to connect with those interested in developing innovative AI systems for software development.

Beyond the Prototype: Using AI to Write High-Quality Code - Josh Albrecht, Imbue