Potpie

Potpie constructs a knowledge graph of the codebase where functions, classes, files, and services become nodes and calls, imports, and dependencies become edges, powering intelligent inference engines for contextual code understanding.

Potpie has developed a SWE-bench agent that generates file diffs resolving real-world GitHub issues, validated by accompanying test suites.

Potpie's Custom Agents allow teams to define specialized SDLC workflows — debugging, testing, code analysis, feature development — each with tailored prompts and evaluation criteria distinct from general-purpose coding assistants.

A quick-start / agent-skills documentation page is reachable at https://docs.potpie.ai/cli/introduction.

Agent-native positioning with a concrete operational path: "The CLI introduction explicitly describes installing skills into agent harnesses like Claude Code or Cursor, providing a concrete operational path for agents.".

Potpie employs a multi-stage pipeline — research, specification, plan, implementation — that surfaces assumptions and scope decisions as a visible contract before code generation begins.

Before asking clarifying questions, Potpie's agents read and index the repository — grepping for existing patterns, tracing wiring of similar features, and verifying where new code must land — producing a verified map of what actually exists in the codebase.

Potpie supports multiple LLM providers, avoiding single-model vendor lock-in.

Potpie agents have web access capability for retrieving external information during tasks.

Potpie implements Semantic Sandboxing — RBAC applied directly to the code knowledge graph — so that sensitive code paths are unreachable as context rather than merely hidden by instruction.

Potpie's context graph is a structured graph of symbols, relationships, and governance rules — not a visual map — enabling fine-grained understanding of codebase architecture including permissions, team ownership, and regulatory obligations.

The entry page was fetched and analyzed for machine-interface signals (title, headings, developer links, keyword probes).

sitemap.xml is reachable and lists site pages.

A documentation surface is reachable at https://docs.potpie.ai/.

A quick-start / agent-skills documentation page is reachable at https://docs.potpie.ai/quickstart.

A quick-start / agent-skills documentation page is reachable at https://docs.potpie.ai/tutorials/explore-your-codebase.

Potpie provides isolated sandbox execution for agent workflows with persistent workspaces that survive session restarts, protecting host environments from agent actions.

Potpie treats agent evaluation as a continuous loop rather than a one-time benchmark, with a methodology designed to generate harder questions, expose gaps, and measure improvement iteratively.

Potpie's evaluation framework tests dependency awareness across heterogeneous service boundaries, using diverse codebases like Apache Airflow to move beyond language-specific optimizations toward broader Repository Intelligence.