feat: add bundled Automating Your Work skill

[just-cameron]

Overlord (agent-c2adbf5c-8419-4211-8cd8-3740db164974)

Summary

Adds a bundled Building Automation skill. It helps an agent extend itself with one-off and recurring workflows built on the Letta Agent SDK.

What the skill explains

• Call the agent itself in a new or saved conversation so the automation can reuse its memory and identity.
• Select a lower-cost model for a non-default automation conversation without changing the agent model used by other conversations.
• Build one-off helpers, reusable commands, scheduled programs, event-driven programs, and longer-running services.
• Combine instructions for judgment, code for fixed work, and agent turns for interpretation and tool use.
• Use temporary worker agents, persistent specialist agents, or additional conversations when separate context, another model, parallel work, or independent review helps.
• Run tools in managed cloud sandboxes, on connected computers in the Letta organization, on the current machine, or through a remote App Server. Connected computers can also form fleets of CI machines, infrastructure nodes, or test machines.
• Choose options for state, tools, permissions, structured results, reporting, and operations.

Examples

references/sdk-recipes.md includes TypeScript examples for:

• Calling an existing agent for a one-off task.
• Creating or resuming a conversation with its own model selection.
• Creating a hidden temporary worker with its own model and read-only toolset.
• Running workers in parallel.
• Saving conversations by resource.
• Scheduling work and handling events.
• Reporting a worker result to a coordinator conversation.
• Discovering online organization computers and selecting one by stable deviceId.

The worker patterns draw from Sarah’s dynamic workflow examples in letta-agent-sdk#261, including parallel audits, fix loops, planning panels, verified research, and isolated sandbox migrations.

Battle testing

Fresh agents applied the skill to three different requests:

• A deterministic Markdown-to-JSON transformation stayed as ordinary TypeScript and did not use the Agent SDK.
• Repeated GitHub triage used the same agent in a lower-cost non-default conversation and did not add worker agents.
• A 40-file migration used temporary workers, separate cloud sandboxes, bounded concurrency, and a second review stage.
• A live SDK 0.6.3 smoke created a hidden agent, selected Haiku on a non-default conversation, cloned letta-ai/letta-agent-sdk into a managed cloud sandbox, used the restricted Read tool, returned the package name, and deleted the temporary agent. The conversation record showed anthropic/claude-haiku-4-5.
• A live client.computers.list({ onlineOnly: true }) call returned online computers with the documented stable selectors.

The first migration test found two gaps: the worker recipe passed a local cwd to a cloud session and used unbounded Promise.all. The recipe now shows repository seeding into /root/workspace, terminateOnClose, bounded concurrency, tool restrictions, and review staging. A fresh agent retested the revised examples and found no material defects.

Validation

• bun run check — all 12 checks pass.
• Agent SDK examples compile against the published @letta-ai/letta-agent-sdk@0.6.3 types.
• bun run build packages skills/building-automation/ with the same files as the source directory.

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.