• Preferred path: find an existing agent/template and run it.
• Fallback path: generate a new agent, publish it, then run it.

• Multiple agents could plausibly match and you cannot rank them confidently.
• Required parameters cannot be inferred from the original request.
• An error or unexpected result requires a user decision.

• After every tool call, present results in markdown tables. Never show raw JSON.
• Present numbered options only when the next action is ambiguous (see Execution principle). When auto-advancing, narrate the action instead (e.g. "No exact match found. Generating a custom agent...").
• Keep tables to 5 rows max per page. Use pagination when more results exist.

nimble_agents_list

Could not connect to Nimble. Please set your

NIMBLE_API_KEY

environment variable and retry.

$ARGUMENTS

• Target domain/URL if provided.
• What fields/records the user expects in output.
• One or two broad keywords for search (e.g. "amazon", "reviews", "linkedin", "products").
• Completeness check: Does the request already specify a website/URL AND what data to extract? If yes, mark intent as "complete" — this allows auto-advancing through later steps without asking for information you already have.

agent-name-a

url

agent-name-b

query

url

agent-name-c

product_url

agent-name-d

url

agent-name-e

url

Omit option 7 ("next page") when current page shows all results (count <= 5 or last page).

If zero results:

1. Search with different keywords
2. Generate a new custom agent

undefined

• User/agent picks an agent number (1-5) → existing-agent path (step 5).
• User/agent picks "search with different keywords" → ask what keywords, then repeat step 3.
• User/agent picks "show next page" → call

  nimble_agents_list

  with

  skip

  incremented by 5, present next page.
• User/agent picks "generate a new custom agent" → proceed to step 6. Only ask for a description if the original request lacks a website/URL or what data to extract.

title

price

rating

1. Run this agent
2. Go back to search results
3. Generate a new agent instead

**5b.** Map values from the original request to input parameters. Infer as much as possible (URLs from mentioned websites, queries from described topics, etc.). Only ask for required params that truly cannot be inferred:

url

query

**5c.** Call `nimble_agents_run` with `agent_name` and complete `params`. Present results:

name

value

If the generated schema looks clearly wrong for the user's intent, present options instead:

1. Publish and run this agent
2. Regenerate with a different description
3. Go back to search existing agents

If `"error"` or `"failed"`:

1. Retry with a modified description
2. Search for an existing agent instead

**6d.** Repeat generate calls with the same `session_id` until `"complete"`.

**6e.** Once generation is complete, call `nimble_agents_publish` with the same `session_id` **before** running the agent. Confirm:

agent-name

If publish fails with a 409 conflict, the agent was already published — proceed to run.

**6f.** Build params from schema, infer values from the original request. Only ask for values that truly cannot be inferred (see step 5b). Call `nimble_agents_run` and present results using the format in step 5c.

Install: `pip install nimble_python`
Set env: `export NIMBLE_API_KEY=<your-key>`

• {agent-name}

  → the

  agent_name

  from the run.

• {url}

  → the URL value passed in

  params

  .

• {field_prints}

  → one

  f"field: {record.get('field')}"

  line per output field from the agent's output schema. Use actual field names, e.g.:
  python

  print(
      f"title: {record.get('title')}, "
      f"price: {record.get('price')}, "
      f"rating: {record.get('rating')}"
  )

• If the agent's input schema has params beyond

  url

  , pass them via

  extra_body

  :
  python

  response = client.extract(
      url="{url}",
      skill="{agent-name}",
      extra_body={"{param}": "{value}", ...},
  )

  Only include non-URL params that were actually used in the run.

Could not connect to Nimble. Please set your

NIMBLE_API_KEY

environment variable and retry. Get a key at app.nimbleway.com → Account Settings → API Keys.

nimble_agents_get

Agent "{name}" was not found. It may have been removed or renamed. Use

nimble_agents_list

to search for available agents.

The agent returned no results. Possible causes:

• The target URL may be unreachable or behind authentication.
• The page structure may have changed since the agent was created.
• Required parameters may be missing — check the agent's input_schema.

processing

processing

Agent generation is taking longer than expected. You can wait or try a simpler prompt.

nimble_agents_publish

nimble_agents_list

• This skill is for agent workflows only — list, get, generate, run, and publish. Do not suggest scheduling, monitoring, automation, cron jobs, or any functionality outside the allowed tools.
• Do not suggest or offer capabilities beyond what the allowed tools provide.
• Never ask for information already present in the original request. Infer URLs, fields, and parameters from context.
• Only present numbered options when the next action is genuinely ambiguous. When auto-advancing, narrate the action.
• Always publish generated agents before running them (step 6e).
• Session IDs must be reused across generate and publish calls within the same flow.
• Present tool call results in tables. Never show raw JSON.
• Adapt table columns to match actual data returned; templates above are examples.
• For the generate path, only ask for extraction description if the original request lacks a target website or what data to extract.