The Coding Agent is the company’s internal agent for repository-scale software maintenance, built on the pi agent harness. It integrates MCP connectors to proprietary codebases and quant-data systems, executes in a sandboxed environment, and ships with cost/latency observability and evaluation pipelines benchmarked against open-source agents.

​
Surfaces

  • Slack — @-mention the agent in a thread with a GitHub link or a request; it checks out the repo and works in the background. Unclear requests get follow-up questions first (clarify mode), and each Slack thread maps to a resumable agent session.
  • TUI — a terminal interface for interactive development.
  • Headless — one-shot and streaming modes for automation:

​
Connectors and tools

​
Sandboxing and guardrails

  • A guardrails extension blocks denied shell patterns and writes to protected paths before any tool executes; rules live in guardrails.json.
  • Extensions are always loaded from the agent’s own repo root — never from a checked-out repository — so cloned code can never inject tools or override guardrails.
  • A read-only plan mode restricts the agent to inspection.

​
Observability and benchmarking

Every session emits JSONL telemetry — turn and tool latency, token usage, cost, and model switches — to .zen/traces/<sessionId>.jsonl. Open-source models are first-class (DeepSeek, Kimi, Qwen, GLM, …; self-hosted vLLM / SGLang / Ollama via env vars), and closed models plug in the same way — so benchmarking cost/latency per task across models, and against open-source agents, is a matter of swapping the model, not the harness.

Agent API

Surfaces, modes, configuration, and the trace schema.