RMT Engineering Logo
Agent and flow builder

AI Agent Builder

Describe the agent in plain English and OptiML writes the first draft: persona, prompts, knowledge scope, tools and guardrails. Then open the canvas and change anything you disagree with.

  • Plain English to working agent
  • 100+ nodes on one canvas
  • Simulate before anyone calls
We reply within one business day. No newsletter.
100+ Workflow nodes
35+ Prebuilt flow templates
5 Agent engine types
One paragraph From description to scaffolded agent

The blank canvas is what stops most teams

A blank canvas, an empty, unconfigured visual workspace, becomes the fundamental starting point to design the AI agent's workflow from scratch as per the preferences of the Customers. Once the Client writes the first draft in their own terms and required workflows, every piece of it is then laid out separately, and none of it applies until the approval from the Client. What the Client accepts becomes a real agent on the canvas that is ready to deliver.

How it fits together

The path a conversation takes through AI Agent Builder

A hybrid agent chaining the FAQ, NLU, flow and prompt engines with fallback, ending in a handover to a human Handover to a human Customer turn FAQ engine deterministic match NLU engine classify and route Flow engine structured steps Prompt engine grounded LLM no match routed not covered Answered — same knowledge, same guardrails, same customer record, whichever engine replied Nothing confident to say → handover, with the conversation attached A hybrid agent chains all four with fallback. Most production agents are hybrid.

Precision first, fluency last. The engine that answers changes; the knowledge scope, the guardrails and the customer record do not.

By the numbers

Few Facts - at a glance

  • 100+

    workflow nodes across logic, data, communication, chat and connectors

  • 35+

    prebuilt flow templates, plus 12+ industry agent templates and 10 vertical packs

  • 5

    agent engine types: prompt, FAQ, NLU, flow and hybrid, on one agent model

The process

How It Works

The four stages, end to end

Four stages, and you can stop at any of them. Some teams describe an agent and publish something close to what came back. Most open the canvas first, correct it, then simulate. Nothing reaches a customer until the version you are looking at on screen is the version you deliberately released, and if that turns out to be the wrong call, the one before it is still sitting there.

  1. Describe it

    Write a paragraph: what the agent should do, which languages it works in, when it should give up and hand over. OptiML drafts the whole configuration from that and lays it out one piece at a time. Accept a piece, edit it, or throw it away. None of it touches the agent until you do.

  2. Open the canvas

    Everything the scaffold wrote sits on the canvas, and the canvas is the agent rather than a picture of it. Drag in nodes. Branch with if/else or switch, set retries and a failure route per node in the property inspector, and drop to a JavaScript node on the day the palette runs out of ideas.

  3. Simulate the change

    Run the new version against synthetic customers and let the evaluation harness score it. The awkward turns are where agents come apart: the caller who interrupts halfway through the verification script, the one who answers a yes-or-no question with a paragraph, the one who changes their mind on the last step and wants to go back two. A simulation finds them first.

  4. Publish, or roll back

    Publishing snapshots the whole behaviour: prompts, model choice, voice settings, tools, guardrails, turn-taking. That single build then runs on voice and on every chat channel. If the change reads badly on live conversations, roll back to an earlier snapshot.

Features

Every capability you need in one module

1. Natural-language agent building

Write a paragraph describing what you want. Something like "a Hindi and English agent for our loan collections team that checks EMI status, takes a promise-to-pay, and transfers anyone who disputes the amount." OptiML scaffolds that into a persona, a system prompt, an engine type, a knowledge scope, tool bindings, escalation rules and guardrails. Nothing is applied until you accept it, piece by piece. Reject the lot and you have lost two minutes.

  • Persona and system prompt, drafted from your own description
  • Guardrails listed one at a time
  • Engine type and knowledge scope chosen and shown, never assumed quietly
  • Tool bindings and escalation rules proposed with the trigger that fires them written out, so you can disagree with it while it is still a proposal

See how knowledge grounding and citations work

Illustrative interface. The description and the scaffold are sample data.

2. Five engines, one agent model

Prompt agents answer from a system prompt through an LLM. FAQ agents match deterministically for high-precision answers. NLU agents classify intent and route. Flow agents run structured step-by-step conversations. Hybrid agents chain all four with fallback, and that is what most production agents end up being.

  • Turns are offered to the engines in order of precision
  • An engine hands on only when it has nothing confident to say
  • What changes is the engine that answers, not the knowledge scope, the guardrails or the customer record
  • The last fallback is a human
Prompt

Prompt agents answer from a system prompt through an LLM. The most open of the five, and the one that most needs grounding and guardrails underneath it.

FAQ

FAQ agents match deterministically for high-precision answers. Where the answer has to be word-for-word, such as a regulated disclosure or a fee schedule, determinism beats fluency.

NLU

NLU agents classify intent and route. This is the engine that replaces a menu tree: the caller says what they want and lands in the right place without pressing anything.

Flow

Flow agents run structured step-by-step conversations. Verification, data capture and anything with a required order belongs here, because a step cannot be skipped by being asked nicely.

Hybrid

Hybrid agents chain all four with fallback, which is what most production agents use. FAQ first for precision, NLU to route, flow for the structured part, and the prompt engine to handle everything that was not anticipated.

3. The visual canvas

A drag-and-drop designer: node palette down one side, property inspector down the other, live execution panel showing what a test run actually did. 100+ nodes covering logic, data, communication, chat and connectors. Branch with if/else, switch, filter or parallel paths. When the logic outgrows the palette there is a JavaScript node and an expression evaluator waiting.

  • Palette, property inspector and live execution panel, side by side
  • If/else, switch, filter and parallel branches
  • Set per node in the inspector: a stored credential, a retry policy with exponential backoff, and a failure route that can transfer to a human
  • JavaScript node and expression evaluator when the palette runs out

The same canvas used for back-office automation

Illustrative interface. Node names and timings are sample data, not a performance claim.

4. Templates and vertical packs

35+ prebuilt flows across seven verticals, plus 12+ industry agent templates you can open and edit. Ten vertical packs go further: one click activates a working set of agents, a knowledge base, compliance templates and workflows configured for banking, insurance, healthcare, retail, travel, telecom, real estate, education, government or HR and IT service desk. All of it editable afterwards.

  • 35+ prebuilt flow templates across seven verticals
  • 12+ industry agent templates you can open and edit
  • A vertical pack activates agents, a knowledge base, compliance templates and workflows together, in one step
  • Nothing a pack installs is locked
Banking Insurance Healthcare Retail Travel Telecom Real estate Education Government HR and IT service desk

Talk to us about the vertical pack for your industry

5. Same agent, every channel

An agent built once runs on voice and on every chat channel. The engine adapts the presentation: quick-reply buttons on WhatsApp become spoken options on a phone call. No second build. No second configuration to drift out of step with the first one.

  • One build for voice and all eight chat channels
  • Presentation adapts per channel, logic does not
  • Nothing to drift, because there is no second copy of the agent sitting somewhere else

See all eight chat channels in one inbox

The same question, two channels

On WhatsApp the agent offers three quick-reply buttons. On the phone the same three options are spoken, and the caller can interrupt to choose. One flow, one knowledge scope, one set of guardrails. The difference is presentation, and the engine decides it.

6. Version, simulate, roll back

Every agent version is a behavioural snapshot: prompts, voice settings, model choice, tools, guardrails, turn-taking. Before a change goes near a customer you can run it against a set of synthetic customers who behave the way awkward callers behave, and the evaluation harness scores the run without anyone having to listen to all of it. If the change still reads badly on live conversations, roll back. A bad edit costs an afternoon, not a week of bad calls.

  • A snapshot captures the whole behaviour, prompt text included
  • Simulation runs before publish, not after
  • The evaluation harness scores the run, so the decision to ship has a number behind it rather than a hunch
  • Rollback to any previous version, without rebuilding

See the evaluation gate a version has to clear

Use Cases

Where AI Agent Builder delivers value

Natural-language agent building

A collections agent written as a paragraph

Trimurti Finance (illustrative), a mid-market lender

Scenario

The collections manager writes a paragraph in her own words. Two languages, an EMI status check, a promise-to-pay, a transfer for anyone who argues about the amount. What comes back is a hybrid agent with a collections knowledge scope, two tool bindings, an escalation rule and five proposed guardrails. She rejects one of them, softens the persona from firm to plainly courteous, and accepts the rest.

Outcome

A first draft she edits rather than authors, and nothing configured that she had not read first.

Canvas, inspector and custom code

The rule the palette could not express

Anchor Assurance (illustrative), a motor insurer

Scenario

Claims routing turns on an internal rule about vehicle age, policy tier and garage network that nobody has ever managed to write down properly. The team builds the conversation from the palette, then drops a JavaScript node in the middle to compute the routing decision. The inspector gives the core-system lookup three retries and a transfer-to-human on failure.

Outcome

The exception lives in one node the team can read, instead of in a prompt that has to be argued with.

Version, simulate, roll back

A prompt change that never reached a customer

Helio Telecom (illustrative)

Scenario

A rewritten system prompt is meant to shorten billing explanations. Simulated against synthetic customers first, it scores badly on the disputed-charge turns, because the shorter answer quietly drops the adjustment window. The version goes back for an edit and a second simulation before anyone publishes it. Nobody outside the team ever hears the short version.

Outcome

The regression turns up in a simulation run rather than on live calls. Either way the previous snapshot is one click behind.

At a glance

Specification

The numbers and limits, without the sales copy

Specification for AI Agent Builder
Specification Detail
Creation paths Natural language 9-step wizard templates vertical packs flow designer REST API
Engine types Prompt, FAQ, NLU, Flow, Hybrid (chained with fallback in a hybrid agent)
Canvas @xyflow/react node palette property inspector live execution panel
Nodes 100+ across logic, data, communication, chat and connectors
Node settings Stored credentials, retry policy with exponential backoff, per-node failure route
Triggers 50+: event webhook email calendar schedule manual
Templates 35+ flow templates 12+ agent templates 10 vertical packs
Custom logic JavaScript code node and expression evaluator
Versioning Behavioural snapshots with synthetic-customer simulation and rollback to any previous version
Channels per build Voice plus eight chat channels from one agent
FAQ

Questions,
answered

What teams ask us before they roll out AI Agent Builder — how it works, what it needs from your side, and what happens when it gets something wrong

Still not sure?

Talk to a specialist and get a straight answer.

Ask our team

For most agents, no. Agent Builder covers the common cases through the plain-English description route and the template library, and the canvas is drag-and-drop with a property inspector rather than a config file to hand-edit. You will want a developer for the JavaScript node, the expression evaluator or the REST API. In practice that is one or two nodes inside an otherwise visual flow.

No. Agent Builder presents the scaffold as a list of proposed pieces, and each one is accepted or edited by you before it exists on the agent. The languages it picked and the escalation trigger it chose are written out rather than left implicit, and the guardrail count on screen is a proposal rather than a setting. Until you accept, the agent is unchanged.

Through nodes on the canvas: HTTP request, SQL query and the connector set, authenticated with a credential stored in your tenant rather than typed into the flow. Each of those nodes in Agent Builder carries its own retry policy and its own failure route, so a slow core-banking call transfers to a human instead of leaving the customer listening to nothing. The tenant itself runs in the region you choose, so nothing about the integration changes where the data sits.

No. An agent built in Agent Builder runs on voice and on all eight chat channels from a single build, and the engine decides the presentation. Three quick-reply buttons on WhatsApp are three spoken options on a call. The logic, the knowledge scope and the guardrails are the same object, so there is no second configuration to drift out of step.

Every Agent Builder version is a behavioural snapshot covering prompts, model choice, voice settings, tools, guardrails and turn-taking, so a bad change is a rollback rather than a rebuild. Before publishing, run the version against synthetic customers and let the evaluation harness score it. In a live conversation a hybrid agent falls through its engines in order of precision and hands over to a human when none of them is confident. The conversation goes with it, so the customer does not start again.

Keep exploring

Additional supporting modules

Native modules that share the same record, knowledge and governance

Platform module

Automation & Workflows

The same canvas, pointed outward: 50+ triggers, API calls, CRM updates, approvals and multi-step journeys.

Explore
Platform module

Knowledge & Grounding

What the agent is allowed to say. Hybrid retrieval scoped per agent, citations, and a groundedness gate.

Explore
Platform module

AI Governance & Guardrails

The gates a new version clears before it reaches a customer, and the caps and audit trail it runs under afterwards.

Explore
Connected solutions

Where AI Agent Builder is used

The Solutions pages that lean on this module, and what it looks like once it is configured for a particular floor, job title or job to be done.

6 solutions built on this module

By industry

Sector floors that run on it

2

By use case

Jobs it is put to work on

4
Browse every solution
Talk to a specialist

Describe your agent on the call itself

Bring the process you would hand to a new starter: a collections call, a KYC check, a renewal conversation. Describe it out loud and we will scaffold the agent from what you said, then open the canvas and change whatever the scaffold got wrong.

  • 30 minutes
  • Your process, scaffolded and edited live
  • Nothing to install

Book your slot

Leave your email and our team will come back to you within one business day.

or reach us directly
Call our team Email us

Your details stay private. We never share them.

This website uses cookies.

Cookies are small text files that allow us to create the best browsing experience for you on our site. By continuing to use this website or clicking "Accept & Close", you are agreeing to our use of cookies. To understand how we use cookies or how to manage them, please see our cookies policy.

Ask OptiML

Powered by RMT Engineering