Autonomous Research Agent

The problem

A search tool bolted to one LLM call fails in three specific ways

Not because the model is weak ; because nothing in that loop is positioned to check the output. The prose is confident either way, which is exactly the problem.

• [S9] no retrieved text
  01

  Citations that resolve to nothing

  The model cites a page it never opened. Nothing captured the text behind that URL at retrieval time, so no later check can confirm the sentence came from it ; or that the page ever said it.

• wire.com news.co press.io → 1 story
  02

  Five links, one story

  A wire report republished across six outlets looks like six sources. Two subdomains of one company look like two companies. Independence was never tested, so agreement is counted where none exists.

• 2021-01 → treated as current
  03

  Numbers that expired

  Without a freshness horizon and a publication date per source, a 2021 roadmap reads exactly like a 2026 one. Undated pages are the worst case: they get silently treated as current.

The answer

Four guardrails, each one a mechanism you can read

All four are enforced after generation, by code rather than by another model. Everything retrieved lands in an EvidenceStore at retrieval time; the gate resolves every claim against that store and nothing else.

When verification fails, the agent gets exactly one bounded repair pass with the specific violations quoted back to it ; and the repair is kept only if it did not increase the issue count. If problems survive that, the report ships with those claims marked ⚠ flagged in the UI. It degrades honestly rather than failing closed or quietly asserting.

How a run works

Seven stages, and the one that costs nothing is the one that decides

An ambiguous question pauses the run and asks ; before any research budget is spent. Everything after that is bounded.

1. clarify

   pauses before budget is spent

2. plan

   sub-questions + freshness horizon

3. research

   ≤ 3 rounds, bounded

4. gap analysis

   what is still thin?

5. report

   streamed, evidence prompt-cached

6. verify

   $0.00 · deterministic

7. repair

   at most one pass

Research stops early on any of three conditions: a round that adds no new sources, a token budget that runs out, or a gap analysis confident enough to stop. Whichever fires is recorded on the run as stopped_because.

What you actually see

The interface shows the evidence, not just the answer

Sources stream in as they are retrieved. The report streams token by token. Every [Sn] marker links to the row it came from, and any claim the gate could not prove carries its reason in words.

Illustration of the product UI. It is populated with the verifier’s own test fixtures from tests/test_verify.py ; including the ir.tesla.com / www.tesla.com pair from TestGuardrail2Independence, which is a real, passing test of the single-source rule. This repository ships no committed example report, so no research findings are shown here.

Where the tokens go

Every ceiling is a constant you can read, and the gate is free

Cost is governed by two things: which model each role gets, and the hard ceilings around the research loop. Both live in researcher/config.py, and both are overridable per run from the request body ; or by env vars on Settings.role_config.

Per-role model routing

The research rounds dominate token volume, so Sonnet 5 carries the bulk and Opus 4.8 is spent only on the report ; a small share of total tokens. There is no cheaper option for the research role: it drives tool use across many turns, and dropping it below Sonnet 5 costs more in wasted rounds than it saves per token. thinking is set explicitly for every role because the API default differs by model.

Hard ceilings per run

Token budget per run

120,000

task_budget_tokens

Research rounds

3

max_rounds

Runner iterations

30

max_runner_iterations

web_search calls

12

web_search_max_uses

web_fetch calls

20

web_fetch_max_uses

Report output cap

64,000

report_max_tokens