Red Yellow BlueContact
AI Agentic Development

Production AI agents for mid-market firms.

Most AI initiatives ship a prototype and stop there. Red Yellow Blue builds agents that run in production, on your infrastructure, with real audit trails, real cost attribution, and real ongoing improvement. Vendor-neutral. Cloud-agnostic. Delivered by the founder directly.

Talk to Tom about your use case

The category

What agentic AI development actually is.

Agentic AI development is the practice of building LLM-powered agents that take multi-step actions on behalf of a user, use tools (APIs, databases, document stores, external systems), maintain state across turns, and operate within governance guardrails. It is distinct from single-shot chatbots and distinct from prompt engineering. It is a full-stack discipline that combines model selection, tool architecture, deterministic pipeline design, security, observability, and operational governance.

An agent that only produces text is not agentic. An agent that produces actions without state, tools, or oversight is not safe. An agent that has all three but sits inside a slide deck is not a business outcome. Agentic development is the specific engineering discipline that turns model capability into shippable, operable, improvable software.

Mid-market firms sit in a difficult zone for this category. The capability is real, the impact is significant, but the skills to deliver it well are scarce, the platforms are moving fast, and the compliance surface is growing (AI management standards, information security frameworks, sector-specific operational risk regimes, and the ordinary data-sovereignty expectations of boards). RYB exists to close that gap.

Six principles

How RYB builds agents.

Six non-negotiables that shape every agent RYB ships. They are the reason a Discovery on Monday can be a production agent by Friday, and the reason those agents run for years afterwards without collapsing under their own complexity.

  1. 01 —

    Deterministic pipelines, creative reasoning

    The LLM reasons and decides what to do. A deterministic pipeline actually does the thing. A job number is never allocated by a model. A fee is never calculated by inference. An email is never sent without a person confirming. Where correctness matters, code owns it; where judgment matters, the model owns it.

  2. 02 —

    Approve-first by default

    Every agent action that affects the world in a lasting way is proposed for approval before it happens. An audit trail lives beside every approval. Trust builds because reversibility is preserved. Where autonomy is safe, it is explicit; where autonomy is dangerous, it is denied.

  3. 03 —

    Per-client cloud environment

    Every client gets their own cloud environment with their own IAM boundary, their own data stores, their own logs. Data never crosses tenants. Costs are precisely attributable. Handover on termination is clean, no matter which cloud the client is on.

  4. 04 —

    Data lives where the client says it lives

    Region and residency are set by the client's sovereignty, regulatory, and contractual requirements. Every deployment is architected around those constraints from day one, not retrofitted later. Documented and audit-visible.

  5. 05 —

    Reusable IP, bespoke delivery

    RYB brings platform IP to every engagement: proven agent runtime patterns, tool-integration frameworks, dashboard scaffolding, deterministic pipeline design. What you get is bespoke to your workflow, built on patterns that are battle-tested across the RYB portfolio.

  6. 06 —

    Productised advisory, not project consulting

    Every capability we ship is designed to be operated for years, not delivered for a quarter. Ongoing improvement runs on a monthly subscription that grows with what you use. New capabilities are scoped and quoted individually. You always see the number before you commit.

The stack

The technology RYB standardises on.

Vendor-neutral does not mean tool-agnostic. RYB has a considered stack that has proven itself across the client portfolio. New tools are added deliberately, not chased. The stack is documented so clients know what they are inheriting.

Inference and reasoning

  • Frontier LLM providers. we work across the leading commercial and hosted model families; model choice is a per-workload decision, not a house rule.
  • Hosted inference platforms. we deploy onto whichever hyperscale AI platform fits the client's cloud footprint, data residency, and cost profile.
  • Task-tuned models where they earn it. small classification models for high-volume routing, frontier models for reasoning, open-weight models where the workload calls for it.

Agent runtime

  • Long-running agents with session state. multi-turn, tool-using agents with persisted context and observability.
  • Event-driven and scheduled workers. agents that wake on a schedule, on a webhook, or on a business event.
  • Model Context Protocol (MCP). the emerging standard tool interface for LLMs; every integration we build exposes an MCP surface where it makes sense.

Tools and integrations

  • Any modern SaaS the client already runs on. CRMs, ERPs, finance and accounting systems, document platforms, marketing automation, industry-specific vertical SaaS.
  • Enterprise document estates. shared document platforms via native APIs and app-only auth, not brittle scraping layers.
  • Legacy and industry systems. older APIs, exported data, and specialised systems handled with the same discipline as modern ones.
  • Custom domain integrations. where no vendor API exists, we build the integration; where one does, we use it properly.

Infrastructure

  • Infrastructure-as-code end to end. every deployment is version-controlled, reproducible, and portable across cloud providers.
  • Client-owned state and storage. data stores, object storage, and logs live in the client's environment under their control.
  • Observability and audit. every agent action is logged, attributable, and reviewable; costs are precisely attributable per client and per workload.

Production examples

Agents in production, right now.

Real systems running for real businesses. Named on the case studies page where the client has consented, and described here specifically enough to be verifiable either way.

RYB internal

Our own agent platform

80+ production agents that run RYB itself: CRM and finance system sync, contract discovery and canonicalisation, meeting synthesis, commitment watchers, engagement enrichment, cashflow monitoring, opportunity detection. Orchestrated by a mission-control frontend. Battle-tested on our own operations before anything ships to a client.

Building consulting

Job intake and fee proposal automation

A multi-channel agent (chat, email, web form) that automates job intake and fee proposal generation for a building consulting firm. Runtime deployed into the client's own cloud environment. Writes directly into their document platform. Deterministic pipeline enforces the rule that the LLM never allocates job numbers, using an atomic counter. Published as a public case study.

Technology distribution

Customer and sales-rep agent runtime

End-customer and sales-rep agents running as a multi-tenant hosted service for a national technology distributor. One runtime, many customer relationships. Integrated with the distributor's product catalogue and pricing systems.

RYB product

Fee-letter generation as a product

A multi-tenant fee-letter generation platform operated by RYB. Exposes an agent-callable API so agents in client environments can create fee letters as tool calls. Enforces money as integer cents, single-source totals, and the three-total separation that makes fee arithmetic auditable.

Project management consulting

Project cost dashboard and CRM-PM integration

A live project cost dashboard for a project-management consulting firm, delivered alongside a broader AI Discovery. Bi-directional integration between the client's PM system and CRM in production for 18+ months. Client-facing project reporting, and the foundation layer for a proprietary contract benchmarking database now under construction.

Enterprise technology distribution

Enterprise integration plus progressive AI layering

Multi-year Enterprise-tier engagement covering deep CRM integration, an internal knowledge bot, and progressive layering of AI capabilities on top of business-critical infrastructure. Reference case for how a large distribution business scales AI adoption without disrupting operations.

Commercial model

Build plus subscribe.

RYB does not sell projects and disappear. The productised advisory model exists because agentic AI is a capability that improves continuously, not a one-shot delivery. Every engagement has two commercial lines that fit together.

Initial build. Scoped, fixed price, typical delivery in 2 to 6 weeks depending on complexity. Anchor-scale builds often bundle into the ongoing subscription rather than invoice separately; Core and Enterprise builds are quoted individually.

Monthly subscription. Covers cloud hosting, platform operations, incremental improvement, monthly steering, and async access. Anchor tier starts at $2,000 (AUD) per month. Core tier engagements land in the $60,000 to $120,000 per year range with a proper vCIO shape and dashboard build. Enterprise engagements start from $200,000 per year with dedicated capacity.

Boundary rule. The subscription keeps everything already built alive and improving in small ways. Anything meaningfully new (a new agent, a new integration, a new job type) gets scoped separately so you always see the number before you commit.

Governance

How safe agents actually get built.

Agent safety is not a policy statement. It is a stack of specific engineering decisions that make the wrong thing hard and the right thing easy.

Human-in-loop by default. Every state-changing action starts as a proposal. The person named as the actor confirms before the action is committed. An audit trail records who confirmed, when, and what happened afterwards.

Deterministic separation. Anything that must be exactly right (a job number, a fee amount, a client id lookup) runs through code, not through the model. The model can propose it, but code owns the truth.

Vendor-neutral posture. RYB is not a reseller for any vendor that ends up in a client architecture. Cloud platforms, model providers, CRM and finance systems, and vertical software are chosen on architectural fit, not on commission schedule. Where a client-preferred vendor changes the picture, RYB says so transparently.

Compliance alignment. For regulated clients, engagements align with the relevant standards for information security (such as ISO/IEC 27001), AI management systems (such as ISO/IEC 42001), operational risk regimes (such as sector-specific prudential standards), and applicable privacy law in the client\'s jurisdiction. RYB is not a certification body; the architecture is designed to make certification tractable for the client.

FAQ

Questions RYB gets often.

What is agentic AI development?

Agentic AI development is the practice of building LLM-powered agents that can take multi-step actions on behalf of a user, use tools (APIs, databases, document stores), maintain state across turns, and operate within governance guardrails. Distinct from single-shot chatbots or prompt engineering, agentic development is a full-stack discipline that combines model selection, tool architecture, deterministic pipeline design, security, observability, and operational governance.

What makes Red Yellow Blue different from a generic AI consultancy?

RYB ships production agents, not slide decks. Every engagement lands in a per-client cloud environment with real infrastructure, real logs, and real audit trails. The founder writes production code directly; there is no analyst layer between the client and the delivery. Vendor-neutral (no reseller commissions), cloud-agnostic, and structured around a productised advisory model so ongoing improvement is affordable at mid-market scale.

Which AI platforms and models does RYB use?

RYB is deliberately platform-neutral. Every engagement is architected around the client's existing environment, security posture, and data locality needs. In practice we work across the major frontier LLM providers, all three major hyperscale cloud AI platforms, and standard agent protocols like Model Context Protocol (MCP). Model and platform selection is a per-engagement decision driven by task fit, latency, cost, and data-residency requirements, not a house preference.

How do you handle data sovereignty and security?

All data, all documents, and all model inference stay in the client's chosen region. Every client gets a dedicated cloud environment with its own IAM boundary. Approve-first is the default: agents propose actions, humans confirm before writes. Deterministic pipelines separate creative reasoning from state-changing actions (numbers, ids, and identifiers are never allocated by an LLM). Full audit trails on every action.

What does a typical AI agentic development engagement cost?

RYB uses a build-plus-subscribe model. Initial builds are scoped and fixed-price. Ongoing operation and incremental improvement runs on a monthly subscription that grows with the client. Anchor-tier engagements start around $2,000 (AUD) per month plus scoped build work. Core-tier engagements sit in the $60,000 to $120,000 per year range, and Enterprise engagements start from $200,000 per year.

Do I need to be a tech company to work with RYB on this?

No. RYB serves mid-market services firms across property, professional services (design, engineering, project management, legal), financial services, healthcare, technology distribution, and building consulting. Clients typically have 14 to 200 staff and want to use AI to grow without linear headcount growth. The methodology adapts to the industry, but the architecture patterns are consistent.

Can you show me production examples?

Yes. RYB Brain is our own operational agent platform running 80+ agents in production. Client engagements in production cover job-intake and fee-proposal automation in building consulting, multi-tenant customer and sales-rep agents in technology distribution, CRM-native AI assistants in professional services, and multi-tenant fee-letter generation as a hosted product. See the published case studies for named examples, and book a call for architecture diagrams and outcomes not on the public page.

Further reading

Related from the RYB library.

Talk to Tom

Bring a real use case. Get a real conversation.

30 minutes with Tom Leyden directly. Bring a use case you would like to explore. We work through the architecture, the technology fit, and whether RYB is the right partner. No sales team. No pressure either way.