Letta’s office hours for July 2, 2026 centered on a naming shift with architectural consequences. Cameron explained that the broader packaged experience delivered by Letta Desktop should now be called Letta Agent, while the underlying harness remains Letta Code. That distinction was presented as a way to better reflect how people actually use the system: not only for software development, but also for operations, knowledge work, research, companionship, and creative play. The talk quickly widened into a product-and-community tour of the week’s most visible experiments.
This guide is part of the Letta Office Hours archive and describes the episode as a historical record rather than a current product specification.
From there, the episode became a survey of how the platform was evolving around that broader agent model. Cameron highlighted the end of the Mod Challenge, the arrival of new model options, a preview of subagent and reflection UI in the app, and ongoing work to make Slack a serious surface for human-agent collaboration. The through-line was consistent: Letta is moving from a code-centric harness toward an orchestration layer for many kinds of agents, with mods, skills, and channels acting as the mechanisms that shape behavior.
Selected chapters
- 00:00 Intro and office hours format
- 00:22 Letta Agent naming and Letta Code distinction
- 01:43 Four major Letta Agent use cases
- 02:53 Mod Challenge wrap-up
- 03:41 Community mod highlights
- 07:34 Sonnet 5 and Fable 5 availability
- 09:23 Letta agents orchestrating Claude Code and Codex
- 10:10 Subagent and reflection panel preview
- 11:49 Slack agents and virtual coworker workflows
- 13:33 Slack run/status UI improvements
- 20:23 Agent SDK, app server, and channels
- 37:56 Organizing complex agent knowledge with skills and MemFS
Subject areas
Naming the product around what people actually do
The renaming discussion was not cosmetic. Cameron argued that “Letta Agent” better fits the lived use cases people describe: coding, operations, knowledge management, research, companionship, and role-play. In the episode’s framing, “Code” still names the core harness, but “Agent” names the broader experience around it. That separation matters because it clarifies which layer is infrastructural and which layer is user-facing.
Mods as a behavior layer
The Mod Challenge served as the best proof that a lot of useful agent behavior can be added without changing the core harness. Cameron walked through examples such as Jukebox, Auto Pivot, Control Room, Environment Compass, Hyper, Oath Keeper, Sprite, ThreadKeeper, and Muscle Memory. The point was less the individual gimmicks and more the pattern: mods can add coordination, guardrails, memory, and quality-of-life features that make agents more capable and easier to trust.
Slack as a collaboration surface
A major product thread was Slack integration. Cameron described Slack as a place where agents can act like coworkers inside the same workflow as humans, and showed work on inline run blocks that expose what an agent is doing while it works. The design goal is transparency: commands, reasoning, and progress should be visible in context, making Slack a practical operational surface for persistent agents rather than just a chat inbox.
Subagents, reflections, and orchestration
The app preview focused on a new panel for subagents and reflection agents. Instead of hiding those processes in popovers, the UI surfaces them in a dedicated side panel so users can inspect what happened, what files changed, and what a reflection concluded. That supports a broader architectural idea: one agent can delegate to many specialized workers, and the user should be able to see and reason about that delegation.
Models, cost, and choosing the right tool
The model discussion was pragmatic rather than hype-driven. Cameron compared Sonnet 5, Fable 5, and GLM-5.2, noting tradeoffs around cost and availability. The recurring recommendation was to let Letta agents orchestrate heavier CLI-based coding sessions when appropriate, because the outer agent can manage the larger task while subagents do the detailed work.
Q&A themes
The Q&A clustered around a few recurring concerns: how mods are scoped, how skills are learned, whether agents can fork themselves for debugging, how local mode interacts with event-driven features, and how to organize knowledge with skills and MemFS. Several questions probed the boundary between mod state and durable memory, and Cameron repeatedly returned to the idea that the harness observes behavior, then distills repeated patterns into reusable structure.
Architectural through-line
The episode’s architecture story is delegation. Letta Agent is presented as a layered system: a harness at the bottom, channels and mods as runtime behavior, subagents and reflection agents as coordination machinery, and skills/MemFS as a way to turn repeated work into durable capability. Slack then becomes the external collaboration surface where this machinery is visible to humans. The result is not just a chatbot or a coding assistant, but an extensible agent workplace.