A research agent that shows its work.

Biotech due diligence normally takes a room full of specialists: legal reviewing the patents, BD sizing the strategy, clinical and preclinical reviewers weighing the data, and executive teams evaluating the evidence and business case. The question I set out to answer was whether a single agent could carry that load, or at least the evidence-gathering underneath it.

The hard part was never the writing. It was trust. A general model writes fluent citations that don’t resolve, and reads a fuzzy sequence off a patent scan then corrects it into the wrong one. In diligence, a wrong PMID or a single wrong base pair is not a rounding error. So the whole build points at one thing: guardrails that make every output checkable.

I named it after my son. Sonny is a multi-agent research agent for biotech that only makes claims it can support and verify.

General models can produce answers that sound authoritative while citing papers that do not exist or trials that are slightly wrong. For diligence work, that is a dangerous failure mode because it looks correct even when it is not.

Sonny addresses this with two checks. First, a structural gate verifies that every citation token resolves to a real retrieved record. Then a verifier checks whether the cited passage actually supports the claim. If the evidence is missing, Sonny abstains instead of inventing a story.

I built Sonny as a TypeScript monorepo around a small, legible core. Every specialist works from the same fixed rubric, and every claim moves through the same five stages before it reaches you. No model composes the rubric.

In use, you ask about a target. Six specialists take it up along fixed axes: target biology, mechanism, indication, clinical precedent, competitive IP, and modality risk. The orchestrator streams the trace as they work: searching ClinicalTrials.gov, fourteen trials found, nine claims verified, one flagged. You read the reasoning as it forms instead of staring at a spinner.

This is where general models are most dangerous: they often turn a blurry sequence into a plausible one, changing a base or two with complete confidence. For sequence work, that is worse than failing loudly.

Sonny treats the scan as the least reliable source. It uses the patent number to retrieve the canonical sequence from public databases and verifies it by exact match. Only if no canonical source exists does it fall back to OCR, where it constrains decoding to valid residues, checks the patent’s stated length, and BLASTs the fragment to identify the likely sequence. Each result includes provenance and a confidence flag.

Specialists plan their own research questions. Four of five were getting dropped. Each round pursued the first open question, then replaced the whole queue with whatever the reflection step suggested next. A ledger merges now instead of replacing, and a question counts as answered only when a grounded claim answers it. The specialist no longer grades its own sufficiency.

So the dossier ships with the questions the specialists could not answer. An absence claim, “no precedent exists”, needs a retrieval audit showing which searches ran. A search that came back genuinely empty counts, because that is what the absence rests on. Hits thrown out by the relevance gate do not.

I route models by job shape: a strong planner for orchestration, cheaper models for extraction, and a verifier from a different family than the writer so it can catch blind spots.

Sonny runs end to end on local models over free public sources, so a full research pass carries no API cost. Ollama by default, or a hosted model if you have one. The engine ships as a library, not a service, and LUMINA embeds it.

The verify-every-claim design costs latency and roughly doubles token spend, but that is the tradeoff for diligence. The evaluation harness compares a raw model, the same model with web search, and Sonny on a small golden set plus a fictional-gene abstention trap. The key metric is verifier performance on claims that already pass the citation gate: supported, unsupported, or overreach. A second set measures whether the specialists dug at all.

I sequence the roadmap so every capability lands with a measurement in front of it. Nothing ships without a number that can catch it regressing.

A trustworthy agent does not come from asking nicely in the system prompt. It comes from building a system that cannot ship an unverified claim. The real shift is from prompting for better behavior to enforcing it in the architecture, and that becomes visible in everything the agent refuses to say.