A document library with a team of AI specialists.

You bring your documents — research, policy, client materials, working drafts. The library keeps them versioned, searchable, and organised. The specialists write from your sources, verify every claim against the evidence, and evaluate the result against your standards.

Work that has been fact-checked and scored, not just generated.

We are in closed beta. To find out more, get in touch.

How it works.

Five specialists work with your library: Collector finds and files source material. Composer writes from your sources. Corroborator verifies claims against the evidence. Critic evaluates work against your standards. Curator keeps the library organised as it grows.

They work together in sequence: gather, write, verify, evaluate, file. Each document is versioned, searchable, and has its own URL and visibility controls.

You can also connect your own AI tools — Claude, Cursor, or similar — and use the library directly as a managed document store. Most users do both.

Multi-agent AI systems that work in production.

What we've found is that prompting alone isn't enough — agents follow structural constraints but unreliably follow instructions.

If your organisation is deploying AI agents and struggling with reliability, hallucination, specification drift, or agents that work in demos but fail in production, we can help. Our approach draws on organisational theory — separation of duties, adversarial review, information compartmentalisation — applied to multi-agent systems.

We offer consulting on AI agent architecture, reliability engineering, and the design of multi-agent workflows that produce trustworthy output. Get in touch.

Observations from running this in production.

We keep a public log of observations from operating this system in production. Read our field notes.

Our approach is grounded in organisational theory — the study of how structure produces reliable collective behaviour from individually unreliable actors, whether human or artificial. Read the underlying theory:

The research core of the Perseverance Composition Engine is available at codeberg.org/leithdocs/persevere.

Plumbing — a typed composition language for agent pipelines.

Get in touch.

Tell us what you're working on. We typically reply within one working day.