Agent Custodian

The workplace for
specialist AI agents

Create, run, and govern AI agents built for your organisation's systems, policies, and institutional knowledge.

Start for free
Agent Custodian networkNetwork Judge · governance onlyFinanceHRLegalResearchOperationsSENTINEL · MEMORY GATEMother BrainCOMPANY-LEVEL

  • Mother Brain

    Company memory that every agent learns from and shares.

  • Data Brain

    Connects agents to your ERP, HR, CRM and data warehouse.

  • Custodian Mesh

    Checks every plan: approve, hold for a person, or block.

  • Sentinel

    Decides which lessons are safe to keep in memory.

  • Network Judge

    Oversees the network while each company's memory stays isolated.

Solutions

Built for every team

Agent Custodian ships purpose-built workflows for the teams that need them most.

  • Finance

    Automate reconciliation, audits, and compliance checks at scale.

    • Reconciliation
    • Audit prep
    • AML / KYC
    Shares company memory

  • HR

    Streamline onboarding, policy lookup, and performance reviews.

    • Onboarding
    • Policy Q&A
    • Reviews
    Shares company memory

  • Legal

    Draft contracts, flag risks, and accelerate due diligence.

    • Contracts
    • Risk flags
    • Due diligence
    Shares company memory

  • Research

    Synthesize literature, generate hypotheses, and track citations.

    • Literature
    • Hypotheses
    • Citations
    Shares company memory

  • Operations

    Automate workflows, monitor SLAs, and surface bottlenecks.

    • Workflows
    • SLAs
    • Bottlenecks
    Shares company memory

  • Healthcare

    Support clinical documentation, triage logic, and coding tasks.

    • Clinical notes
    • Triage
    • Coding
    Shares company memory

How it works

From idea to running agent in three steps

  1. Create an agent

    Pick a template or write your own identity. Name it, set its goal, and add constraints.

  2. Ask it anything

    Type a question. The agent recalls company knowledge, writes a plan, and runs a safety check automatically.

  3. Review and evolve

    See every step in a clear timeline. Hold for human review. Accept suggested updates when they improve the agent.

Inside every run

Five steps. Always visible.

Every answer shows its working: what the agent remembered, how it reasoned, the plan it wrote and the safety check it passed. When a plan needs a person, the run waits.

  1. Memory Recalls the company knowledge that matters
  2. Reasoning Thinks the request through
  3. Plan Writes the actions it will take
  4. Safety check Checks the plan against your policies
  5. Answer Replies, or waits for a person
A completed run: memory, reasoning, action plan, an approved safety check and the final answer A completed run: memory, reasoning, action plan, an approved safety check and the final answer
Completed
A run waiting for review: the safety check flagged a wire transfer and asks a person to approve or reject A run waiting for review: the safety check flagged a wire transfer and asks a person to approve or reject
Needs review

Platform

Everything you need to ship

One platform. From prototype to production, with enterprise guardrails baked in.

  • Multi-Agent Orchestration

    Compose specialized agents into pipelines that handle complex, multi-step reasoning.

  • Universal Connectors

    Plug into 200+ enterprise systems via REST, GraphQL, or native SDKs in minutes.

  • Network Judge™

    Every output is evaluated by a consensus judge network before it leaves the system.

  • Trace & Replay

    Full observability into every token, tool call, and decision with replay capability.

  • Enterprise Security

    SOC 2 Type II, SSO/SAML, private VPC deployment, and end-to-end encryption.

  • One-Click Deploy

    Ship agents to production with blue/green deployments and automatic rollbacks.

  • Agent overview with a suggested update banner, identity, goal and constraints Agent overview with a suggested update banner, identity, goal and constraints

    Agents that suggest their own updates

    After enough runs, an agent proposes a sharper identity. You review the diff and decide.

  • Connections page listing REST connectors and an MCP tool server, all reachable Connections page listing REST connectors and an MCP tool server, all reachable

    Your systems, registered once

    REST connectors and MCP tool servers, tested and ready for every agent that needs them.

Network Judge™

Hallucinations don't reach production.

Every output is evaluated by a consensus ensemble of specialized judge models before it reaches your users. Accuracy, safety and policy are all gated automatically.

  • Ensemble voting across 5+ judge models
  • Custom rubrics and domain policies
  • Automatic retry on low-confidence outputs
  • Full audit trail for every decision
  • run_7f2a PASS
  • run_3b9c RETRY
  • run_1d4e PASS
  • run_8a2f BLOCK
  • run_5c7d PASS

How a decision flows

Every output passes the judge before it leaves. Verdicts are written to an append-only audit chain.

One view of the whole network

Organisations, agents and decisions, live, at millisecond resolution.

Network map: a platform core connected to organisations, next to a live decision feed Network map: a platform core connected to organisations, next to a live decision feed

Sentinel

Only safe lessons become memory.

Agents learn from every run. Before a lesson is written to your organisation's memory, Sentinel judges it against the compliance packs your residency requires, so personal data, injected instructions and made-up figures never become how you do things.

  • Residency picks the checks

    Kenya applies the Data Protection Act, the EU and UK apply GDPR, and every region applies the CAI and UN packs. Your data stays where it is.

  • Core checks run at full strength

    Prompt injection, toxicity, redlines, personal data and memory quality are always on and cannot be turned down.

  • Fails closed

    If Sentinel cannot be reached, the lesson is frozen. Nothing is stored silently.

  • Every judgment is on record

    Each verdict keeps its region, the packs applied and the reasons, ready for review.

LessonALLOWInstitutional memoryLessonFREEZEHeld for reviewLessonSEVERRefusedSentinelDPA KEGDPRCAIUNJudgment log · region, packs, reasons

Features

Built for how organisations actually use AI

  • Version control for agents

    Every edit creates a snapshot. Roll back to any prior identity in one click. No prompts lost, no surprises.

  • Safety on every run

    A built-in safety check reviews each plan before execution. Approve, reject, or let it proceed automatically.

  • Company memory

    Agents draw from institutional knowledge on every run. The right context, every time, without you pasting it in.

  • Start from templates

    Finance, Research, HR: pre-built identities that work out of the box and customise in minutes.

  • Connect your systems

    Register REST connectors and MCP tool servers. Agents know which systems they can reach, ready for when you wire them up.

  • Agents that evolve

    Suggest an update; the agent drafts its own improved identity for your review. Accept or decline. You stay in control.

Of specialist agent
3 types
Per run, always visible
5 steps
Human-controlled approvals
100%
Surprises (by design)
0

FAQ

Common questions

What is Agent Custodian?

Agent Custodian is a workplace for specialist AI agents. You create agents from templates or your own identity, run them against your company knowledge, and govern every step with safety checks and human review.

How does the Network Judge work?

Before an output leaves the system, an ensemble of specialised judge models scores it for accuracy, safety and policy. Low-confidence outputs are retried automatically, and anything that fails is blocked and held for review.

Can I connect my own LLM?

Yes. Agents can use the hosted models we support or a model endpoint you provide, and every run still goes through the same safety checks and judging.

Is my data used for training?

No. Your prompts, knowledge and run data are never used to train shared models. Your workspace data stays isolated to your organisation.

What connectors are available?

You can register REST connectors and MCP tool servers today, with a growing library of ready-made connectors for common enterprise systems.

How do I migrate from another platform?

Most teams start by recreating their prompts as agent identities and registering their existing APIs as connectors. Our team can help plan larger migrations.

What is the uptime SLA?

Enterprise plans include a contractual uptime SLA. Contact sales for the details that apply to your deployment.

Can agents run on my own infrastructure?

Yes. Enterprise customers can deploy into a private VPC so that agents, knowledge and run data stay inside their own environment.

Is there a free tier?

Yes. The Starter plan is free and includes 3 agents and 1,000 runs per month, with no credit card required.

How do I get enterprise pricing?

Use the contact form below or email our sales team, and we will put together a plan that fits your security and scale requirements.

Get started

Ready to ship your first production agent?

Start free. No credit card required. Deploy in minutes.

Start building free Talk to sales