Coco
Use cases

AI drug discovery

How Coco governs an agent drafting regulatory answers and lab tool submissions in drug discovery and product development, and what we have not yet measured.

Decide each action a drug discovery agent takes against your own policy, before it runs.

Agents in drug discovery and product development now submit modelling jobs, look up compounds, and draft answers to regulatory questions. Coco decides each of those actions against the organisation's own policy before it runs, and keeps the record.

Keep in mind: No drug discovery team runs Coco today.


The problem in drug discovery work

A lab agent calls tools built for people. A structure job sent through a connector is wrong when the agent skipped a conversion the interface applies, or guessed a field the interface measures. Nothing fails, and the result looks normal.

Regulatory work carries the same exposure. Each label and claim has to stand against the regulator's text in each market, and expert reviewers are the bottleneck. A system prompt asks the model to cite its sources. It doesn't stop an invented citation.

How Coco governs a drug discovery agent

Coco sits between the model and its tools, and decides before an action runs, in code the model doesn't control. A tool reached through an MCP server, or a send to an outside service, is a tool call like any other.

Control

Rules decide first, from the organisation's policy and against facts from its systems. An optional classifier only tightens a verdict. A rule error is a deny, and with ungoverned: DENY a tool no contract governs doesn't run.

Two contracts (behaviour contracts, each written from a rule) were drafted for a lab agent's job submissions. Neither is in use. One denies a structure model job when a conversion the model's spec declares hasn't run. The other denies a job when the agent guessed a field derived from the structure instead of measuring it. In the contract format:

# An example, drafted and not in use
id: lab.structure_job_inputs
status: draft
action: lab.submit_job
checks:
  - id: declared_conversion_ran
    check: 'state.job.conversion_required == "" or state.job.conversion_applied == true'
    on_fail: DENY
    reason: The model's spec declares an input conversion, and it has not run on this input
  - id: derived_fields_measured
    check: 'state.job.fields_guessed == 0'
    on_fail: DENY
    reason: The agent guessed a field derived from the structure instead of measuring it

Intervention

Every action ends in one of four verdicts: AUTO_EXECUTE (run), OBSERVE (run and flag), ESCALATE (pause for a person), or DENY (deny with the reason). In this design, Coco denies a skipped conversion before any job exists, and sends a draft for a market the reference data doesn't hold to a regulatory reviewer. No contract applies either verdict in this domain today.

Evidence

Every decision leaves a receipt in a hash-chained ledger, and each rule traces to its clause. Receipts aren't signed. For regulatory answers, the check looks up every citation again in the retrieved material.

Policy from the organisation's own documents

The document pipeline reads SOPs and policies into AI-ready context and draft contracts, each citing its clause. A contract runs in testing first, recording what it would decide, and goes live when a person promotes it. The pipeline hasn't been tested on a pharmaceutical SOP.

Public regulation takes a second route. Coco built a regulatory reference database, FDA first, from 58 open sources, and fingerprinted each download. It holds 24,089 substance rules, 102,992 approved and registered products, 39,218 approved US drug labels among them, and 12,711 sections of law and guidance. A row the schema fails to read cleanly goes to a person, and a market the database doesn't hold reads as not covered, never as clean.

The gaps are published: 6,819 substance rules carry no identifier, 411 restricted rules state a condition with no readable number, 175 carry neither limit nor condition, and 3,877 enforcement events carry no date.

Keep in mind: The database isn't part of the product today.

How Coco fits the drug discovery workflow

Coco sits at each tool call. CocoGuard is middleware in coco-framework, LangChain, or LangGraph. The hosted gate serves any other runtime over HTTP. The Claude Code hook checks Claude Code's tool calls, MCP tools included.

The regulatory check runs as a demonstration and returns no verdict. Three lookups run with no model, then a model drafts the answer. The demonstration didn't keep compound data inside the organisation: an unknown structure went to a public chemistry service, and a third party hosted the drafting model. Coco offers to run both inside your environment, and hasn't done so yet.

What the guardrails stop

There's no drug discovery benchmark. The nearest measurement is the regulatory demonstration of 17 October 2026:

  • Citations: On two regulatory questions (n=2), every citation traced to retrieved material: 31 of 31 and 35 of 35, none invented. This measures tracing, not correctness. Where coverage is missing, the answer says so and cites nothing.

The engine's figures come from payments and injection benchmarks, run on 29 September 2026:

  • Payments: On a corpus of 73,356 payment transfer calls built from 4,371 human-written attacks, rules let 0 of 15,550 refused transfers run and held 0 of 57,806 permitted ones, at 0.27 ms per decision at the median.
  • AgentDojo v1.2.2: Of 588 planted-instruction attacks, rules alone prevented 467. Rules plus classifier prevented 582, with 18 of 97 legitimate tasks held for a person.

To reproduce, run:

# from a coco-sdk checkout; the third line needs the classifier running
coco eval vault --arms rules                     # run 20260929-153604
coco eval agentdojo --arms none,rules            # run 20260929-185238
coco eval agentdojo --arms judge,hybrid --question v2 --classifier <classifier URL>   # run 20260929-184422

# from the regulatory database export, available on request
python3 -c "import csv; r=list(csv.DictReader(open('_holdings.csv', encoding='utf-8-sig'))); print(len(r), *(sum(int(x[k]) for x in r) for k in ('rules', 'products', 'sections')))"
cat _data_quality.csv
python3 -c "import json, sys; print(json.load(open(sys.argv[1]))['check'])" <answer record>.json

What Coco can't promise yet

  • No lab or regulatory benchmark exists, and the classifier has never been measured on lab or regulatory traffic.
  • No contract reads the regulatory database, and the lab contracts are unused drafts.
  • The database holds none of 21 CFR Parts 11, 58, 211, and 312, no clinical trial or pharmacovigilance rules, and no EMA, TGA, or ASEAN rules. Coco claims no compliance with any of them.
  • No integration exists with an ELN, a LIMS, or lab instruments.

From our team

"An agent in regulated product development earns its place when every action it takes and every sentence it drafts traces to a source a reviewer opens. Coco checks each action against the organisation's own rules before it runs, keeps the record, and leaves the regulatory decision with the person who signs it." Khoa Pham, co-founder and CTO, CocoAI

Working with us

If you run agents in drug discovery or regulated product development, contact us at trustcoco.ai. We start with your own regulatory questions answered with sources, a contract on your agent's tool calls, or a test against your labelled cases before anything goes live.


On this page