Potpie

Potpie is an AI coding platform that takes a fundamentally different approach to understanding your codebase. Instead of treating source files as flat text to feed into a language model, it constructs a structured knowledge graph — where functions, classes, files, and services become nodes, and calls, imports, ownership, and dependencies become edges. This graph powers specialized AI agents that can debug issues, generate code, write tests, and develop features with an awareness of how your system actually works.

How the Pipeline Works

Potpie's workflow follows a deliberate multi-stage pipeline: research, specification, plan, and implementation. During the research phase, the agent reads and indexes the repository — grepping for existing patterns, tracing how similar features are wired, checking which primitives are already in use, and mapping where new code needs to land. According to Potpie's engineering blog, "anything the codebase can answer, the agent should answer before it asks the developer."

The specification stage turns this research into a visible contract that teams can review and challenge. If the proposed scope is incorrect, behavioral assumptions are flawed, or key constraints are missed, those issues surface before a single line of code is written. Only after approval does the plan stage convert the specification into concrete build instructions.

Knowledge Graph and Semantic Sandboxing

The context graph underpinning Potpie's agents is described by the company as "not a visual map" but "a structured one — a graph of symbols, relationships, and governance rules all living together." This architecture enables what Potpie calls Semantic Sandboxing: applying role-based access control directly to the graph so that sensitive code paths — payment gateway integrations, fraud detection logic, authentication internals — simply do not exist as reachable context for agents operating under restricted permissions. For that session, "the secret simply does not exist as reachable context," which Potpie positions as a fundamentally more trustworthy architecture than instructing an agent to withhold information it can already see.

Evaluation and Benchmarking

Potpie publishes its own evaluation framework and has built a SWE-bench agent designed to generate file diffs that resolve real-world GitHub issues, validated by accompanying test suites. The company's testing spans diverse codebases including Apache Airflow, chosen for its multi-service distributed system context, complex workflow DAGs, and database migrations. The evaluation methodology emphasizes dependency awareness — understanding how changing a utility function in one module breaks schema validation in another — and Integration Fidelity for code generation agents. Potpie frames evaluation as a continuous loop rather than a one-time certification: generate harder questions, expose gaps, fix the agent, measure again.

Platform Capabilities

The platform supports multiple LLM providers and has recently added web access capabilities for agents. Custom Agents allow teams to define specialized workflows — debugging agents that follow different patterns than feature-development agents, each with tailored prompts, tool access, and evaluation criteria. The platform operates within its own sandbox, providing workspace persistence so that agent sessions survive restarts.

Potpie competes in the AI Agent Development space. Its graph-based approach and structured pipeline differentiate it from single-prompt coding tools, while its sandbox architecture and semantic access controls aim to make it suitable for teams working on compliance-sensitive codebases. Comparable tools in this category include Openclaw and Genspark.ai.