Skip to content
radicalsoft

Reference project

AI Agent Platform

Orchestrate agents that use tools, memory and knowledge, with human approval where it counts.

Discuss a similar solutionAI Engineering
Laptop screen showing lines of programming code

Overview

What it is and the problem it addresses.

A reference platform for building, running and governing AI agents inside a company, with the observability and controls operations teams need.

The problem

Teams can prototype an agent in a day, but running agents safely requires orchestration, permissions, evaluation and a way to stop actions that need a person.

  • Agent orchestration
  • Tool calling and MCP integrations
  • Persistent memory
  • RAG and knowledge retrieval
  • Workflow automation
  • Human approval controls

Main features

What the concept includes.

  • Orchestration

    Define agents, roles, budgets and handoffs, with retries and timeouts managed centrally.

  • Tools and MCP

    A registry of typed tools and MCP servers with per-agent permissions.

  • Memory

    Short and long-term memory scoped by user, team and task.

  • Knowledge retrieval

    Permission-aware retrieval over approved sources, with citations.

  • Workflows

    Compose triggers, agents, conditions and actions in a visual builder.

  • Approvals and audit

    Approval queues for sensitive actions and a full trace of every run.

Architecture overview

How the pieces fit together.

  1. 01

    Control plane

    Agent definitions, permissions, budgets and versions.

  2. 02

    Runtime

    Workers that execute plans, call tools and persist state.

  3. 03

    Tool and knowledge layer

    MCP servers, internal APIs and retrieval services.

  4. 04

    Observability

    Tracing, evaluation and cost reporting.

Example workflows

How it is used.

Support triage

  1. 1Email arrives
  2. 2Agent classifies and drafts a reply
  3. 3Refund above threshold goes to approval
  4. 4Reply sent and logged

Research brief

  1. 1Analyst asks a question
  2. 2Research agent gathers sources
  3. 3Review agent checks citations
  4. 4Brief delivered for sign-off

Target users

  • Operations teams
  • Support organisations
  • Engineering and data teams

Potential integrations

  • MCP servers
  • CRM and ticketing
  • Databases and search
  • Messaging platforms

Security considerations

  • Scoped credentials per agent
  • Approval gates for irreversible actions
  • Prompt-injection defences on tool output
  • Full audit trail
See all projects

Want something like AI Agent Platform?

Tell us about your goals and constraints. We will outline how a similar system could be built for you.

Discuss a similar solutionAll projects