Patty Code
The Korean-first AI coding agent
A controllable coding agent that converses in your terminal, runs tools, and stays with the task until it’s done.
Patty Code is a coding agent that converses in your terminal, executes tools, and sees tasks through to completion. Korean prompts are first-class, and you set the ceiling on agent autonomy.
Models and tools are swappable. No vendor lock-in, and a cache-first system-prompt architecture lowers the cost of repeated sessions through byte-stable prefixes. Hierarchical memory (PATTY.md/AGENTS.md, @imports, remember tool) keeps project context alive across sessions.
Validated internally first; public availability is decided after that validation completes. Contact us about organization-level pilots today.
Technology behind this product
Evidence
Architecture
Patty Code’s execution model is a cycle: observe → propose → approve → execute → record. The agent reads files and proposes commands; policy auto-executes what it allows while the rest waits for human approval.
Every tool call is journaled. Reopen a session to replay what was done and why; revert problematic changes per commit.
Benchmarks
We evaluate with our own suite: an e2e task corpus stratified across 8 workload classes, a SWE-bench Verified subset (50 instances), and a context-maintenance A/B harness.
All figures pass grader-validation discipline (fail-on-seed/pass-on-reference) before publication, always with hardware and configuration context stated.