Work Agent

Agent Creator

Create global agents in Treasure Work — either on-demand (invoked from the UI, chat, or @mention) or scheduled (cron-driven). Both share one AGENTS.md format and lifecycle. Setting a schedule: field is the single switch that turns an agent into a scheduled one.

Agent Location

Each agent is its own directory with a single AGENTS.md:

• Global (the only scope this skill creates): ~/.treasure-work/agents/{name}/AGENTS.md
• Workspace (read-only via MCP; edit through the AgentSettings UI): {workspace}/agents/{name}/AGENTS.md

The directory is the agent's working directory at run time, so the body can reference relative paths and the agent can create supporting files as needed. No subdirectories are required.

Workflow

CRITICAL: Never just create and stop. Always run the agent and iterate until it works.

1. Capture intent — what the agent should do, how often (if scheduled), what tools / skills it needs, where results should go.
   • Ask the user for notification channels (notify.on_success / notify.on_failure) before writing. Never assume a Slack channel — always confirm.
2. Create — call agent_create with name, body (the system prompt), and the initial frontmatter (description, icon, skills, schedule if recurring, etc.). New agents default to status: draft so scheduled ones don't fire until activated.
3. Test run — call agent_run_now to trigger an immediate run. It returns a chatId.
4. Inspect output — call chat_read with that chatId. While in_progress: true the run is still streaming — poll until it goes false. Read the last assistant turn to confirm the agent did what was intended.
5. Review — load the agent-review skill for a structure + quality audit.
6. Iterate — call agent_update to patch any field (body, skills, allowed_tools, schedule, …). Re-run with agent_run_now.
7. Activate — once the test run is clean, agent_update with status: active. Scheduled agents start firing on their cron; on-demand agents become first-class in the UI list.

Steps 3–5 are mandatory — an agent is not done until it has been run and reviewed at least once.

AGENTS.md Anatomy

YAML frontmatter + Markdown body. The body is the system prompt the agent sees on every run — write it in the second person as instructions, not as a task description.

---
name: daily-sales-report
display_name: Daily Sales Report
description: Fetch sales data, analyze trends, and post to Slack
icon: 📊
color: blue
status: draft            # draft | active | paused | resting
schedule: "0 9 * * 1-5"  # 5-field cron, ≥5-minute granularity; omit for on-demand
catch_up: false          # true = catch up one missed run after app restart
model: claude-sonnet-4-6 # optional per-agent model override
max_turns: 20
timeout: 600             # seconds, max 3600
autonomous: false        # true = Supervisor Agent auto-continues across turns
skills:
  - sql-skills:trino
allowed_tools:
  - Bash
  - Write
  - slack_post_message
  - slack_upload_file
notify:
  on_success: slack:channel-name   # or slack:dm
  on_failure: slack:channel-name   # or slack:dm
---

You are the Daily Sales Report agent. Each weekday morning:

1. Query yesterday's sales (`td_query` against the orders table) and gather revenue, order count, top products.
2. Compare with the previous run by looking at the last successful chat (use `chat_read` if needed).
3. Post a Slack summary; if revenue < $10K, flag the dip in the message.
4. If anything fails, retry once then surface the error via `notify.on_failure`.

Frontmatter field reference

Do NOT write Slack channel names or notification targets in the body. They belong in notify.* and are injected into the prompt at run time. Hard-coding them in the body causes drift when the frontmatter is updated.

Scheduled vs. on-demand

Permissions

Use short tool names in allowed_tools — both SDK built-ins (Bash, Write, Edit, WebFetch, WebSearch, NotebookEdit) and Treasure Work MCP tools without the mcp__work__ prefix (e.g. slack_post_message, td_query, render_chart).

Always allowed (no entry needed): Read, Glob, Grep, ToolSearch, Skill, Task*, TodoWrite, mcp__td-docs__*, plus the planning / MCP-resource tools.

Notification targets

Use slack:channel-name for a Slack channel, or slack:dm for the user's DM. Always use slack:dm exactly — not "direct message", "DM", or other variants.

MCP Tools

Important behaviors

• Mutations are global-only. agent_create / agent_update / agent_delete operate on ~/.treasure-work/agents/ only. Workspace agents are read-only via MCP — edit them through the AgentSettings UI.
• No separate reload tool. Every mutation broadcasts AGENTS_CHANGED, so the agent list and the open AgentSettings page refresh automatically.
• No separate validate tool. Validation happens on agent_create / agent_update; the error message lists each invalid field.
• Unknown frontmatter keys are preserved. If the file has keys outside the current schema (e.g. fields written by a future version), the writer leaves them alone.

What's gone from the Schedule era

The Schedule-Task model has been retired in the Agents architecture. When authoring or reviewing an agent through this skill, do not write or call the fields and tools below. (Other skill docs in this plugin may still mention some of them — that's stale guidance pending its own update; the agent runtime no longer supports them.)

• TASK.md — the body lives directly in AGENTS.md.
• schedule.yaml — the frontmatter lives directly in AGENTS.md.
• Required subdirs scripts/ / reference/ / data/ / results/ — none. Create files only as needed.
• results/{run_id}/output.md + metadata.json — every run produces a chat instead; read it with chat_read.
• enabled: true / false — replaced by status: active | paused.
• permissions.allow — flattened to allowed_tools at the top level.
• context.max_turns / context.timeout / context.autonomous — flattened to top-level fields.
• cron: — renamed to schedule:.
• Storage roots ~/.tdx/schedule-tasks/ and {workspace}/schedules/ — replaced by ~/.treasure-work/agents/ and {workspace}/agents/.
• schedule_* MCP tools (create_schedule, schedule_list, schedule_get, schedule_validate, schedule_reload, schedule_run, schedule_results, schedule_enable / schedule_disable, schedule_delete, schedule_kill) — all replaced by the agent_* surface above.