Joey Wang
About Joey Wang
Joey Wang built Spectre, the collaborative cloud agent platform inside Harvey, the legal AI company. Spectre runs coding agents such as Claude Code, Codex and Cursor in the cloud with Harvey's secrets, real stack and shared rules, and it has grown from engineering into data questions and marketing work in Slack. In an AI Engineer talk, he describes what broke after 10,000 agent PRs.
Notable Positions
On the Harness as the Product
Harvey doesn't compete with coding agents, it runs them. A laptop handles five to ten agents; beyond that a company needs a cloud layer with secrets, sandboxes, one review queue and fair capacity. See Agent Harness and Cloud Agents.
On Owning What Happens After the Code
A follow-up loop reads CI failures and pushes fixes, and approved PRs can auto-merge. In October four of five agent PRs merged into Harvey's backend had been through the loop.
On Capping What Anyone Can Start
An alert storm launched about 900 agent tasks in four hours, and one engineer's 950-task CLI burst filled the cluster. Harvey responded with alert deduplication, a fair-share launch queue, and automations that switch off after five consecutive failures and can carry budgets or expiry dates.
Key Quotes
- "the agents were the easy part." (on building Spectre)
- "We don't compete with them. We run them." (on Claude Code and Cursor)
- "That's the difference between an agent that writes code and an agent that finishes the job." (on the follow-up loop)
Related Reading
- Cloud Agents - Coding agents running in remote environments
- Agent Harness - The infrastructure around the model