Capabilities that support persistent work
- Longer instruction following lets models work for hours or days when paired with suitable context compression and compaction. This is Lee's qualitative account, not a measured reliability rate.
6:14 - Sub-agents provide fresh working context for expensive tool calls or research, then report results to the main agent. Delegation reduces the main conversation's context burden.
6:41 - Better tool use allows agents to work with shell commands, files, a full Linux computer, and demo recording.
7:157:32 - Computer use lets agents click through websites and SaaS products that lack an API or MCP server. Files give the agent a searchable record of information it may need later.
7:48 - Demonstrated workflows can become reusable skills, memories, or routines.
8:17
What gets stored
| Artifact | Purpose | Source |
|---|---|---|
| User profile | Include stable preferences, such as lowercase messaging, in every prompt | 8:39 |
| Working log | Record work over time so questions such as “What did I do last week?” can be answered by searching files | 8:39 9:04 |
| Skill | A Markdown file encoding the process and success criteria for a task | 9:18 |
| Plugin | Install tools through skills or servers; Grokbot is the example in the transcript | 9:44 |
| Routine | Run a prompt on a schedule, analogous to a cron job | 9:44 |