Deep Dive · XiaoHu Explains

Give every department its own AI, and it gets harder to use: Sierra collapsed four into one — and it now writes 70% of the company's code

A retrospective: Pinecone has run 75,000+ sessions since launch, serves 600+ employees, and connects to 37 internal systems through an MCP Gateway.

Quick Take
  • Sierra's six-person AI acceleration team spent a few months merging four role-specific agents — for support, data analysis, engineering, and sales (PINE, Pinewood, Pinecone, Reggie Jr) — into one unified agent, now called Pinecone.
  • Since launching in March, Pinecone has run over 75,000 sessions, serves 600+ employees, and now originates 70% of the company's code pull requests (PRs).
  • Pinecone connects to 37 internal systems through a self-built MCP Gateway: it inherits each employee's own permissions, runs a policy check on every call, isolates customer data, and leaves an audit trail.
  • The team found more value in having the agent stay with a project and proactively step in when a task is ready, rather than only answering on demand — though most sessions today still start with a human.
  • The team admits its current metrics — session counts, PR counts — only measure activity, and it hasn't yet found a good way to measure whether outcomes are actually getting better.
This piece is adapted from Sierra's official blog, describing the company's own internal AI rollout in its own words. Session counts, user numbers, PR share, and other figures are self-reported by Sierra and have not been independently verified.
01Starting Point

Why everyone wants to clone their best engineer

Sierra's six-person AI acceleration team spent the past six months building internal agents for every department at the AI customer-service company, and wrote up what they learned along the way.

This past March, Sierra merged its four role-specific agents — for support, data analysis, engineering, and sales — into a single unified agent named Pinecone. Four months later, it has run over 75,000 sessions, serves 600+ employees, and now originates 70% of the company's code pull requests.

Through its self-built MCP Gateway, Pinecone connects to 37 internal systems — the widest-reaching, most deeply integrated internal agent the company has run so far.

A 1968 study found that the very best software engineers vastly out-produce average ones. For the next fifty years, companies had exactly one answer to that: try to hire those people.

This past January, Sierra's engineering team came back from break energized by frontier-model progress and started running agents in parallel using git worktree, Claude Code, and Codex. On some tasks, output jumped to 5x.

What is git worktree

A built-in Git feature that lets the same repository open several independent working directories at once, so multiple agents can run different tasks in parallel without stepping on each other. Think of it like partitioning one house into several separate rooms — everyone works in their own room, and no one messes up anyone else's stuff.

If agents can 5x an engineer's output within a month, what happens when you bring that acceleration to the whole company? That question is what led to the six-person AI acceleration team — and this piece is their retrospective on what they did and what went wrong along the way.

5x
Output multiple on some tasks after the engineering team started running agents in parallel with git worktree, Claude Code, and Codex in January
6
Size of the AI acceleration team Sierra assembled
02Reorg

Give every department its own AI, and it gets harder — not easier