Research and agent columns

Research columns

A research column browses the web to answer a question about each row, for example “Does {Company} use Kubernetes? Cite a source.” Unlike an AI column, which reasons over data already in the row, it searches and fetches pages.

Config:

Field	Meaning
prompt	The question, with {Column} placeholders
input_columns	Columns to expose to the prompt
max_steps	Search/fetch steps, 1 to 6 (hard cap 6)
output_format	text or json

Bounds that apply regardless of config: 5 search results per step, 1,500 characters of page text per fetch, and a per-cell dollar budget (RESEARCH_CELL_BUDGET_USD, default $0.05) of which 40% is reserved for the final synthesis call so the budget is a true ceiling.

The result carries answer, citations, cost_usd, steps_used and stopped_reason in the cell metadata. Citations are validated against the set of URLs the harness actually fetched; anything else is dropped.

Two execution paths exist: the default ReAct loop that works with any provider, and a native Claude tool-use loop enabled by RESEARCH_NATIVE_TOOLS=1 when Anthropic is the default provider. Both use in-house search and fetch tools behind the SSRF guard and the prompt-injection fence. The native path degrades to the default loop on any failure, so a cell never crashes. An offline evaluation gate compares the two paths on fixtures and fails the build if the native path regresses on correctness or drops below 90% citation presence.

Agent columns

Where a waterfall runs a fixed provider order, an agent column is goal-directed. Each step it re-plans with the same cost-aware planner, picks the next provider, observes the result, reroutes on rate limits instead of failing, and stops when the goal is met, the budget is spent, or the tools are exhausted.

Config:

{
  "goal": "Find a verified work email for the CTO",
  "tools": ["crosslinked", "hunter_io", "apollo_io"],
  "policy": { "max_steps": 6, "max_cost_usd": 0.10 }
}

tools defaults to the target field’s default waterfall; max_steps defaults to 6 and max_cost_usd to $0.10. The per-cell budget is the smaller of the policy cap and the workbook’s remaining ceiling.

Every cell stores a trace you can audit:

{
  "goal": "Find a verified work email for the CTO",
  "steps": [
    { "step": 1, "provider": "crosslinked", "success": true, "value": "Jane Doe — CTO", "cost": 0, "reason": "free first" },
    { "step": 2, "provider": "hunter_io", "success": true, "value": "jane@…", "cost": 0.04, "reason": "best hit-rate per $" }
  ],
  "outcome": "found",
  "spent": 0.04
}

Outcomes are found, budget or exhausted. Traces are stored per workspace under row-level security. Chat can add an agent column for you with the add_agent_column action after confirmation.