Agentic systems
We build custom multi-agent superintelligence (SI) systems on your documents, records and data. Each agent has one job, and an audit step checks every citation against its source before an answer reaches you.
What we build
An agentic system is an SI system that plans its own steps, calls tools and acts on what they return. We build multi-agent systems where a planner splits the work and specialist agents each work one part.
Research and analysis agents
A planner breaks a question into parts and gives each part to a specialist agent. Their findings are audited and combined into one answer with sources. Suited to legal research, due diligence and pricing analysis.
Document and regulation extraction
Agents read contracts, filings, rulebooks and regulation, and return structured records. Each field links to the passage it came from.
Document and regulation intelligenceRetrieval with checked citations
Question answering over your documents through retrieval-augmented generation (RAG). The system retrieves passages from your documents and writes its answer from them. Each citation in an answer must match a passage the system actually retrieved, or the claim is removed.
Internal assistants
Assistants that work inside your systems with each user's own permissions, and keep a log of every action they take.
Agentic securityMCP servers for business tools
MCP servers that give SI assistants governed tool access to your business tools, under rules you set for tools, accounts and actions. Every call is recorded. Our Glesso service gives SI assistants governed access to business tools in production.
Discuss an MCP serverAgents over Rillor data
Our market and compute pricing data connect as tools, next to your own data. Each result keeps its source.
Market data
How a multi-agent system fits together
Every Rillor agent system has the same parts: a planner and specialist agents, the orchestrator-worker pattern, with an audit step before any answer. Each part has one job, and every step is logged so a run can be traced and replayed.
- Planner
- Reads the question, splits it into parts and assigns each part to a specialist agent.
- Specialist agents
- Each works one part, with its own instructions and only the tools that part needs.
- Tools and data
- Retrieval over your documents, your databases, Rillor data and APIs, reached through governed tool access. Every result is logged as evidence.
- Audit step
- Compares every citation in the findings with the logged evidence. A claim without matching evidence goes back for rework or is removed.
- Answer
- Delivered with its citations, each one checked against the source passage.
- Models
- Model families are chosen per project and versions are pinned, so every run can be replayed.
When one agent is enough
One agent is enough when each step depends on the one before it. Split across agents, such a task is broken into artificial parts, and the effort goes to coordination. Splitting helps when a question breaks into parts that can be worked separately, such as evidence gathered from several sources, which is the work our systems are built for.
Running in production
Grellum runs this design in full. Glesso runs its governed tool access.
Grellum
Agentic legal research. In deep mode, a planner assigns each question to four to six specialist agents, and an audit step reviews their work. Every citation must match retrieved evidence.
Every citation checked against the evidence it came from
Glesso
Governed access to business tools for SI assistants. Each customer's credentials are encrypted and used only for that customer's own requests.
Every call runs on its own customer's credentials
Both run on the Rillor platform.
See all systemsTested, then deployed privately
A system goes live once it passes tests built from your own material, in an environment you control.
Evaluation before go-live
We test each system on questions with known answers drawn from your material, plus out-of-scope and adversarial questions. We measure answer quality, citation accuracy and failure cases for each model family and version, and log every run so it can be replayed.
SI evaluationDeployment on private SI cloud
Agent systems run in a dedicated environment on your servers, in your colocation space, or on GPU capacity Rillor sources for you. Your documents, prompts and logs stay in that environment.
Private SI cloud
How a build runs
Four stages. You review the result of each one before the next begins.
Scope
We map the workflow, the sources and the decisions the system supports, and agree what a correct answer looks like.
Build
We build the agents, tools and retrieval over your data, with every step logged.
Evaluate
We run the agreed tests and failure cases, fix what fails and report the results.
Deploy
We deploy into your environment and operate the system, or hand it to your team with runbooks.
Common questions
What is an agentic system?
An SI system that plans its own steps, calls tools and acts on what they return. Some run as one agent. Others split the work between a planner and specialist agents, each with only the tools its part of the task needs.
When do you use more than one agent?
When a question breaks into parts that can be worked separately, which is the work every system we build is made for. A planner divides each question into parts, and each specialist agent works one part with its own instructions and only the tools that part needs, so each agent has one job. How well a design works depends on the task, the tools, the retrieval behind them and the checks on what they produce.
Which models do you use?
Model families chosen per project, with versions pinned, so every run can be replayed. Evaluation measures answer quality, citation accuracy and failure cases for each model family and version.
Where does it run?
On a private SI cloud: a dedicated environment on your servers, in your colocation space, or on GPU capacity Rillor sources for you. Your documents, prompts and logs stay in that environment.
How is it tested?
On questions with known answers drawn from your material, plus out-of-scope and adversarial questions. We measure answer quality, citation accuracy and failure cases, log every run so it can be replayed, and the system goes live once it passes.
Do you build MCP servers?
Yes. We build MCP servers that give SI assistants governed tool access to your business tools, under rules you set for tools, accounts and actions. Every call is recorded.
Research
The newest articles from our Insights library, each with a full research report and its sources.
- Permission-aware retrieval for internal assistantsRAG access control for internal assistants: how retrieval returns only what the asking user may read, from per-document filters to delegated tokens, and what published leaks show.Agentic systems7 min readPDF report
- Agentic RAG and how it differs from standard RAGAgentic RAG vs RAG: standard RAG retrieves once and answers, while agentic RAG lets the model plan searches, grade what it found and search again, at a higher cost.Agentic systems7 min readPDF report
- How reliable are LLM agents on repeated runsBenchmarks from 2024 to 2026 show how often LLM agents succeed on every repeat of one task, and how far a single run moves a score.Agentic systems7 min readPDF report
Notices
SI outputs. Superintelligence (SI) outputs can be wrong. Important decisions should include human review.
Bring us a workflow and its data.
Tell us the workflow, the sources it draws on and the decisions it supports, and we scope the agents, tools and tests around them.