Research Operator

A research operator is a thin loop around general tools. The methodology should live in skills, the math in libraries, and the case-study logic in project code.

The Problem

Hard-coding every possible research move into an agent framework creates a brittle demo. Giving the model unrestricted shell and file access creates an unauditable production risk. The useful middle ground is a narrow operator: expose a small set of typed tools, require skill discovery before implementation, sandbox writes, and stop with an evidence-backed recommendation.

The Pattern

WRONG

prompt = f"""
You are a quant researcher. Use any files and commands you need.
Try improving the case study and tell me what worked.
Task: {task}
"""
print(llm(prompt))

CORRECT

TOOLS = {
    "list_skills": tool_list_skills,
    "read_skill": tool_read_skill,
    "read_file": sandbox_read_file,
    "query_registry": readonly_sql_query,
    "read_parquet": sandbox_read_parquet,
    "run_bash": sandbox_run_bash,
    "edit_file": sandbox_edit_file,
    "done": finish_with_summary,
}

SYSTEM_PROMPT = """
Before implementation, call list_skills and read the relevant skills.
State the hypothesis, cite the skills used, run the smallest valid experiment,
and finish with an honest recommendation.
"""

state = AgentState(task=task)
while not state.quality_gates.get("done"):
    action = model_choose_action(SYSTEM_PROMPT, state)
    result = TOOLS[action.name](**action.args)
    state.tool_trace.append(log_call(action, result))
    state.checkpoint(run_dir / "state.json")

Operator Contract

• Tool surface stays small: read, query, inspect, execute, edit, skill lookup, done
• Writes go to a sandbox or explicit output directory
• Registry and source repos are read-only unless the task explicitly requires edits
• The agent must read relevant skills before changing code or running experiments
• Final output states hypothesis, method, artifacts, skill usage, metrics, and recommendation

Guardrails

• Framework-first design - if the operator owns methodology, skills become decorative
• Unbounded shell - command execution needs cwd, timeout, and write sandbox controls
• No negative-result path - a valid run may conclude that the proposed change is worse
• Missing skill audit - record which skills were read and whether they were followed

Checklist

• Tool schemas and runtime controls are separate from the LLM
• The loop checkpoints state after every tool call
• The prompt requires skill discovery before implementation
• Writes are redirected to a sandbox or declared output directory
• The final summary includes evidence, artifacts, and an honest recommendation