Journey: Agents
Build the agents described in 02-plan.md. Order: MCP servers first (probe must pass before any agent references them) → agents → smoke runs.
When to use
- Dispatched by
journey when current_stage: agents.
- Directly via
/butterbase-skills:journey-agents.
- Skipped (annotated
(n/a)) if the plan lists no agents.
Preflight
If docs/butterbase/03-preflight.md is missing, older than 24 hours, or 00-state.md has app_id: null, invoke butterbase-skills:journey-preflight first. Wait for it to return successfully before proceeding.
Inputs
docs/butterbase/02-plan.md — the Agents section (names, purpose, tool list, visibility, MCP servers needed).docs/butterbase/00-state.md — for app_id.- Any handcrafted spec files under
agents/ in the project (preferred — version-controlled and survives clones).
Procedure
Refresh docs. Call butterbase_docs with topic: "agents". If the cache (docs/butterbase/03b-docs-cache.md) already covers agents, skip.
Register MCP servers first. For each MCP server in the plan:
- Print:
"Registering MCP server '<label>' at <url> (transport=<sse|http|streamable_http>). Proceed?". Wait for yes.
- Call the MCP-servers registration route (or the dashboard's MCP Servers page if a wrapping MCP tool is not yet available — check
butterbase_docs topic agents for the current preferred path).
- Wait for probe result. If
status: unhealthy, surface the error and ask the user to fix (URL, auth header, transport mismatch) before continuing. Do not create agents that reference an unhealthy server.
Build each agent. For each agent in the plan, in order:
a. Print: "About to build agent: <name> (visibility=<v>, tools=<n builtin / m mcp / k function>). Proceed?". Wait for yes.
b. Invoke butterbase-skills:agents via the Skill tool with: agent name, intended behaviour, plan-derived tool list, model preference, visibility/limits. The wrapped skill drafts the graph_spec, writes it to agents/<name>.json in the project repo, and calls validate_agent_spec.
c. If validation fails, surface the Zod issues and loop the agents skill to fix.
d. Safety gate. If visibility != 'private' AND any node can reach a write tool (insert_row, update_row, delete_row, write_storage, or any read_write-mode MCP/function tool), explicitly ask: "This agent is reachable by <visibility> callers and can write data. Set safety_acknowledged=true? (yes / change to private / reduce tools)". Do not proceed silently.
e. create_agent with the validated spec.
f. Smoke: invoke_agent with a representative input. Poll get_agent_run until terminal. If the run errors, debug per the agents skill's debugging procedure; otherwise show the user the final output.
g. Append one line per agent to docs/butterbase/04-build-log.md:
<ISO timestamp> agents create_agent <agent-name> ok (run smoke=<run_id>)
Persist specs to the repo. Confirm agents/<name>.json files are tracked locally. They will be carried by butterbase repo push (run by the templates stage if the user is publishing, or any time the user pushes a snapshot). This is the only way a clone recipient can recreate the agents — the agents table is not part of clone replay.
Tick state. Mark - [x] agents in 00-state.md, set current_stage to the next unchecked stage.
Return to journey orchestrator (or ask "Continue to the next stage? (yes/no)").
Outputs
- One or more agents created against the live app.
- One or more
agents/<name>.json spec files in the project repo.
- One line per agent in
docs/butterbase/04-build-log.md.
Anti-patterns
- ❌ Skipping
validate_agent_spec. Bad specs surface as opaque runtime errors after the agent is live.
- ❌ Creating a
public write-capable agent without rate limits or a daily_budget_usd. The runtime requires safety_acknowledged, but you should also pick hourly per-IP and per-app caps.
- ❌ Hard-coding secrets in
system_prompt or args_template. Read them from ctx.env inside a function tool.
- ❌ Forgetting that the
agents table is not replayed on clone. Always commit the spec JSON and document re-import (butterbase agents create -f agents/<name>.json) in the README.
- ❌ Smoking with
visibility: 'private' and calling it good for a public agent. After the smoke, change visibility and re-test with an unauthenticated curl to confirm the public path works.
1---2name: journey-agents3description: Use as the agents build stage of the Butterbase journey. Implements the Agents section of 02-plan.md by delegating to the `agents` skill for each agent. Registers any required MCP servers, validates each graph_spec, creates the agent, and smokes it via invoke_agent. Skipped if the plan has no agents.4---5
6# Journey: Agents
7
8Build the agents described in `02-plan.md`. Order: MCP servers first (probe must pass before any agent references them) → agents → smoke runs.
9
10## When to use
11
12- Dispatched by `journey` when `current_stage: agents`.
13- Directly via `/butterbase-skills:journey-agents`.
14- Skipped (annotated `(n/a)`) if the plan lists no agents.
15
16## Preflight
17
18If `docs/butterbase/03-preflight.md` is missing, older than 24 hours, or `00-state.md` has `app_id: null`, invoke `butterbase-skills:journey-preflight` first. Wait for it to return successfully before proceeding.
19
20## Inputs
21
22- `docs/butterbase/02-plan.md` — the Agents section (names, purpose, tool list, visibility, MCP servers needed).
23- `docs/butterbase/00-state.md` — for `app_id`.
24- Any handcrafted spec files under `agents/` in the project (preferred — version-controlled and survives clones).
25
26## Procedure
27
280. **Refresh docs.** Call `butterbase_docs` with `topic: "agents"`. If the cache (`docs/butterbase/03b-docs-cache.md`) already covers agents, skip.
29
301. **Register MCP servers first.** For each MCP server in the plan:
31 - Print: `"Registering MCP server '<label>' at <url> (transport=<sse|http|streamable_http>). Proceed?"`. Wait for `yes`.
32 - Call the MCP-servers registration route (or the dashboard's MCP Servers page if a wrapping MCP tool is not yet available — check `butterbase_docs` topic `agents` for the current preferred path).
33 - Wait for probe result. If `status: unhealthy`, surface the error and ask the user to fix (URL, auth header, transport mismatch) before continuing. Do not create agents that reference an unhealthy server.
34
352. **Build each agent.** For each agent in the plan, in order:
36 a. Print: `"About to build agent: <name> (visibility=<v>, tools=<n builtin / m mcp / k function>). Proceed?"`. Wait for `yes`.
37 b. Invoke `butterbase-skills:agents` via the Skill tool with: agent name, intended behaviour, plan-derived tool list, model preference, visibility/limits. The wrapped skill drafts the `graph_spec`, writes it to `agents/<name>.json` in the project repo, and calls `validate_agent_spec`.
38 c. If validation fails, surface the Zod issues and loop the agents skill to fix.
39 d. **Safety gate.** If `visibility != 'private'` AND any node can reach a write tool (`insert_row`, `update_row`, `delete_row`, `write_storage`, or any `read_write`-mode MCP/function tool), explicitly ask: `"This agent is reachable by <visibility> callers and can write data. Set safety_acknowledged=true? (yes / change to private / reduce tools)"`. Do not proceed silently.
40 e. `create_agent` with the validated spec.
41 f. Smoke: `invoke_agent` with a representative input. Poll `get_agent_run` until terminal. If the run errors, debug per the `agents` skill's debugging procedure; otherwise show the user the final output.
42 g. Append one line per agent to `docs/butterbase/04-build-log.md`:
43 `<ISO timestamp> agents create_agent <agent-name> ok (run smoke=<run_id>)`
44
453. **Persist specs to the repo.** Confirm `agents/<name>.json` files are tracked locally. They will be carried by `butterbase repo push` (run by the `templates` stage if the user is publishing, or any time the user pushes a snapshot). This is the only way a clone recipient can recreate the agents — the `agents` table is **not** part of clone replay.
46
474. **Tick state.** Mark `- [x] agents` in `00-state.md`, set `current_stage` to the next unchecked stage.
48
495. Return to `journey` orchestrator (or ask `"Continue to the next stage? (yes/no)"`).
50
51## Outputs
52
53- One or more agents created against the live app.
54- One or more `agents/<name>.json` spec files in the project repo.
55- One line per agent in `docs/butterbase/04-build-log.md`.
56
57## Anti-patterns
58
59- ❌ Skipping `validate_agent_spec`. Bad specs surface as opaque runtime errors after the agent is live.
60- ❌ Creating a `public` write-capable agent without rate limits or a `daily_budget_usd`. The runtime requires `safety_acknowledged`, but you should also pick hourly per-IP and per-app caps.
61- ❌ Hard-coding secrets in `system_prompt` or `args_template`. Read them from `ctx.env` inside a function tool.
62- ❌ Forgetting that the `agents` table is not replayed on clone. Always commit the spec JSON and document re-import (`butterbase agents create -f agents/<name>.json`) in the README.
63- ❌ Smoking with `visibility: 'private'` and calling it good for a `public` agent. After the smoke, change visibility and re-test with an unauthenticated curl to confirm the public path works.
butterbase-ai