Siqi Chen is a software founder known for Runway, a financial planning platform, and public tools for coding agents. His projects address task completion, model delegation and clearer writing. He also publishes results from internal evaluations of AI models for financial workflows.[2][3][4][5][6]
Start with Taskmaster →
Career
Chen’s speaker biography describes leadership roles at Sandbox VR, Postmates, Hey and Zynga before Runway. The common problem in his current financial work is making a company’s operations understandable through a shared model. Runway Financial operates cfo.ai, where its Ari product connects financial records such as ledger, bank and payroll data and keeps a company model current. Chen’s public agent tools sit alongside that commercial financial software.[2][7][8]
Contributions
Keeping agents working toward the actual goal
Taskmaster handles a familiar failure in coding assistants: a turn ends after partial progress, with a polished summary that sounds like completion. It watches for an explicit session-specific completion token and otherwise injects a continuation prompt into the same session. Arbitrage addresses how to divide the work. Its coordinator writes a specification and acceptance criteria, delegates implementation to Codex, then reviews the returned changes and handles visual validation. Both tools encode an operating procedure around a model rather than claiming to improve the model’s underlying intelligence.[4][3]
Editing for a human reader
Humanizer packages a writing-revision procedure as an agent skill. It uses patterns documented by Wikipedia editors to identify inflated framing, stock phrasing and other structures that can make generated prose less useful. The intended outcome is clearer writing while preserving meaning, rather than evasion of AI detectors. As with Chen’s coding tools, the contribution is a reusable procedure that a person can inspect and adapt.[5]
Tools, writing & talks
Tools
Start here
Taskmaster
github.comA completion guard that checks for an explicit done signal and continues the same agent session when that signal is missing.
Start with the philosophy and core contract to understand what the guard can enforce. Running it requires a supported coding-agent session; read the stop and continuation behavior before adopting the wrapper.
Repository owner[4]Original source- Arbitrage
github.comA skill that separates planning and review from delegated implementation and checks the result against written acceptance criteria.
Repository owner[3]Original source - Humanizer
github.comAn agent skill for revising generated prose while preserving its meaning and removing recognizable writing habits.
Repository owner[5]Original source
Writing, research & talks
Siqi Chen reruns internal evals on Gemini 4 Argon: cost efficiency comparable to Opus 5.5 at introductory pricing but about 2X slower, while gpt-6.1-sol is roughly 3x more cost efficient than either.
Catalog workOriginal source