AI infrastructure · multi-agent systems R&D
A multi-agent environment where the rules are code
In four days we built and open-sourced (MIT) an agent environment where code enforces the rules: cryptography, a PBFT-style agreement protocol, reproducible experiments.
- Status
- Research project
- Timeline
- 4 days
- Stack
- TypeScript / Node.js
- Standard-library cryptography: Ed25519, X25519, AES-256-GCM
- PBFT-style agreement protocol for a fixed set of nodes
- Event sourcing, hash-chained JSONL, deterministic RNG
- LLM adapters: OpenAI-compatible, Anthropic, Ollama
- Docker, GitHub Actions, GHCR
The task
Most multi-agent LLM systems rest on prompt-level rules the model can break. Agents overspend tokens and money, agreement depends on one honest coordinator, and experiments with agent behaviour cannot be repeated or checked by an outsider. We set out to move those rules from prompts into executable code.
What we built
- Built a four-layer runtime on bare Node.js with zero runtime dependencies: agents, stateful links and invariants enforced by code, down to a ban on negative balances.
- Implemented cryptography and a PBFT-style agreement protocol for a fixed set of nodes: Ed25519 identities, encrypted channels, commit certificates and leader change on failure.
- Added a double-entry resource economy with two-sided escrow and atomic settlement; LLM calls are metered inside it, with three provider adapters behind one interface.
- Built a deterministic experiment lab: event sourcing, hash-chained logs, exact replay, sha256 attestation and a read-only observer with API and UI, dependency-free.
- Released under MIT at github.com/StudyLabPro/agent-native-universe, with a tagged release, CI/CD, a container image in a public registry and a hardened runtime profile.
Outcome
- same hash
- a repeated run and a replay of the log give a byte-identical final event hash; anyone can check it on the MIT-licensed code
- 187 of 187
- engine tests passing; master branch, commit 625b81d, as of 05.10.2026
- 4 days
- from an empty repository to a public release: about 20.6K lines of TypeScript, CI/CD and an attested experiment; counted from git history
Services
More cases
Marketplace e-commerce · Pricing
PilotRule-based repricing: a pilot whose contract terms are enforced in code
Competitor prices tracked per SKU, a price computed by rules inside an agreed corridor and applied only after confirmation. Pilot: 10 SKUs, setup within 14 days.
- 10 SKUs · 2 scenarios
- pilot scope; the SKU cap and both scenarios are enforced in code and verified by contract tests, so the cap cannot be exceeded
- up to 14 days
- contractual setup window from client inputs to launch, then a 1-month pilot; the service tracks the dates and closes access when the period ends
Manufacturing · Building materials
MVP in operationPanel specs from CAD drawings in seconds, not hours
Reads DWG/DXF layouts and builds the Excel spec: 10–30 seconds instead of 2–15 hours by hand. All 784 panels of one real project matched the manual spec line by line.
- 784/784 and 189/189
- panels on one project and wall panels on another matched the manual specs line by line, size groups included; area within 0.001 m²
- 2–15 h → 10–30 s
- to produce a project spec: manual work vs app processing, measured on real projects
Marketplace e-commerce · Sales analytics
In productionProfit per SKU across three marketplaces, checked against manual figures
Five seller accounts on WB, Ozon and Yandex Market merged into one SKU-level P&L. Variance from the client’s manual figures on a control SKU for one month: 0.00%.
- 0.00%
- variance vs the client’s manual benchmark across units, revenue, COGS, profit and margin for a control SKU over a month; the threshold was 1%
- 1.18M rows
- of sales across 5 accounts and 12 months processed; 10.7% were ordered and settled in different months
Estimate the payback of a similar process
AVA lets you work out the economics of your process on your own figures, with no form and no price from us. It shows what the manual work costs, what cash effect automation would give and what budget would pay back. If the effect does not even cover running costs, it says so.
Estimate the paybackDiscuss a similar problem
Describe the process and the data you have. Before any work starts we will tell you what can be measured in money, what cannot, and whether it is worth doing.