Jason Liu
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Jason Liu is an engineer at OpenAI who works on OpenAI Codex, presenting practical workflows and design patterns for AI agent productivity in a workshop context. 0:57
Workflow Philosophy
Liu's overarching thesis is that the highest-leverage AI work involves building shared infrastructure and skills, not maximizing personal usage. He argues that "building skills that teammates use is more valuable than maximizing personal token usage." 5:02 This extends to a preference for organizing all long-running AI agent work around a personal monorepo as the single managed project, with other code saved elsewhere. 19:15
He also advocates for letting agents improve themselves over time: allowing agent skills to self-update when they learn something new leads to continuous improvement. 15:03 Relatedly, he recommends delegating prompt writing for automations to the model itself rather than hand-crafting prompts, because model-generated prompts are more in-distribution. 44:45
Context and Memory Management
Liu uses and advocates several Codex-specific features for managing persistent context:
- Compaction: He uses context compaction and considers it effective enough that the old advice to start a new thread after 20 messages is no longer valid. 6:52 3:30
- Memory Vault: He emphasizes investing in a personal memory store as a foundational practice. 4:44
- Pinned Threads: He treats pinned threads with heartbeat automations as persistent teammates — a core design pattern for ongoing agent work — recommending users pin and rename threads to project IDs. 6:08 7:00
Agent Orchestration Patterns
Liu describes inter-agent communication in Codex: agent threads can use list-thread and send-message tools to communicate with each other, enabling manager-level orchestration without human intervention. 6:09
His highest-value pinned thread pattern is what he calls the chief of staff thread — a single thread that checks all connectors and surfaces the most important information daily. 43:05
Input Modalities and Tooling
Liu advocates for dictation as a productivity multiplier, claiming you speak roughly three times faster than you type, making voice input materially more efficient for interacting with AI. 8:11
He is an enthusiastic proponent of Appshots, calling them his "favorite feature of all time" and explaining they are superior to screenshots because they capture the full accessibility tree including element IDs, enabling single-hop tool calls rather than multi-step visual parsing. 16:25
He also advocates for Computer Use, describing it as the first time in a long time he "really sort of felt the AGI of being at work." 16:58
Reasoning Budget Guidance
Liu cautions against defaulting to maximum thinking budget for all tasks, noting that lower reasoning modes remain capable and are sufficient for most automation work. 1:13:19