The platform underneath the roster

Agents are only as trustworthy as the system they run on. Legare Global agents run on a purpose-built agentic operating system: it assigns their work, holds their limits, stops them at sign-off gates, remembers everything, and writes everything down. This page describes the seven parts that matter.

Orchestration

One kernel, many roles

Agents do not freelance. The kernel assigns work, sequences it across roles, and holds each agent inside its remit. When the Contract Analyst finishes a mark-up, the kernel routes it to the right lawyer, logs the hand-off, and starts the clock on review. Cross-role work, such as a supplier contract touching both Legal Ops and Procurement, is coordinated by the kernel, never negotiated between agents.

ROUTING: ROLE-AWARE, REMIT-CHECKED
HAND-OFFS: LOGGED WITH TIMESTAMPS
CONFLICTS: ESCALATED TO A HUMAN

Sign-off gates

Consequential actions wait for a person

Every role carries a schedule of gated actions: sending external correspondence, releasing a report, raising a purchase order. When an agent reaches a gate, the work stops and waits in a sign-off queue for a named approver. The gate records who approved, when, and what they saw at the time. Gates are configured per role at deployment and can only be widened by someone with the authority to do so, which is itself a gated action.

APPROVER: NAMED HUMAN, NEVER A GROUP ALIAS
RECORD: WHO, WHEN, WHAT THEY SAW
CHANGES: GATE EDITS ARE THEMSELVES GATED

Decision log

Every decision, written down

The decision log is the kernel’s memory and your evidence. Every action an agent takes, every draft it produces, every approval and every refusal is written to an append-only log. Entries cannot be edited or deleted, by anyone, including us. Your auditors, your risk team and your regulator (should they ask) read the same record your operations team does.

STORAGE: APPEND-ONLY, TAMPER-EVIDENT
ACCESS: ROLE-BASED, ITSELF LOGGED
EXPORT: YOURS, ON DEMAND, STANDARD FORMATS

The brain

Memory like an employee

A chat window forgets you between sessions. An agent runs on a layered memory: working memory for the task in hand, episodic memory of everything that has happened, semantic memory that maps your firm the way synapses map a brain, procedural memory for how things are done here, and prospective memory for what is due next, all driven by one executive loop. If a server fails mid-task, the agent restores itself from its episodic memory and continues as if nothing happened; your team never sees the seam. And because everything flows through the agent, what your firm knows stops walking out the door when people do.

MEMORY: FIVE LAYERS, ONE EXECUTIVE LOOP
RECOVERY: MID-TASK, INVISIBLE TO YOUR TEAM
KNOWLEDGE: STAYS WHEN STAFF LEAVE

The engine

The best model for every task, every time

No single AI model is best at everything, so the platform refuses to pick one. Each piece of work is routed to the frontier model that does that work best, under one memory and one set of gates. That is why a first draft arrives built to be signed rather than iterated: the document that takes rounds of prompting in a chat window comes back once, researched and structured. Agents deliberately take the time a task needs, and they never stop taking it: every thirty minutes, around the clock, each one checks its position description and picks up the next piece of work.

MODELS: BEST FRONTIER MODEL PER TASK
CADENCE: NEW WORK EVERY 30 MINUTES, 24/7
DRAFTS: BUILT TO SIGN, NOT TO ITERATE

Security

Engineered for firms that cannot get this wrong

The platform we run on was built for financial services first and is aligned with SOC 2, ISO 27001 and GDPR, with the current evidence pack available at any briefing. What is already contractual: data processing agreements at every layer, with you, with our hosting, and enterprise agreements with the model providers, so your data never trains anyone’s model. Agent skills are developed in house and never downloaded from the internet, which closes the door the open-source harnesses leave wide open. Where client data is stored is agreed at deployment, and backups run every hour.

STANDARDS: ALIGNED WITH SOC 2, ISO 27001 AND GDPR
YOUR DATA: DPAS AT EVERY LAYER, TRAINS NOTHING
SKILLS: BUILT IN HOUSE, NEVER DOWNLOADED
BACKUPS: EVERY HOUR

Compounding

Every week it works here, it works better here

Corrections are permanent: tell an agent once and the fix is written into the role for good. Your firm’s document history is ingested in days, so agents arrive already knowing how things are done here. And through an opt-in skills network, when any agent on the platform masters a new skill, your bench can gain it too; the skill travels, your data never does. The result is a workforce that compounds: sharper every month, with none of the ramp-up, and none of the forgetting, that comes with human turnover.

LEARNING: CORRECT ONCE, PERMANENT
ONBOARDING: YOUR DOCUMENT HISTORY IN DAYS
NETWORK: SKILLS SHARED, DATA NEVER

Book a demo

Tell us the role you would otherwise hire for. We will show you an agent at work on it in a sample firm: hand it a task and watch the sign-off gate hold its first draft. No deck, no obligation.

Book a demo Browse example roles

KERNEL v1.0 · NO DECK, NO OBLIGATION