AI Agent Development Australia | Enterprise AI Agents | ComplxAI
AI agent development

Custom AI agent development in Australia

ComplxAI is an Australian team of AI agent developers. Our custom AI agent development covers the whole runtime, from the typed tools an agent acts through to the memory that carries a job to completion and the approval steps that stop it for a person when the consequences warrant it. We design it, build it and deploy it on AWS.

  • Agentic workflows built for Modedu, Luxpip and Billy
  • Modedu doubled revenue in six months; Luxpip cut manual workload by 75%
  • Every stage priced before it begins; the IP is yours from day one
Definitions first

What an AI agent actually is

Strip away the marketing and an AI agent is four things wired together: a model, the tools it may call, memory, and a control loop that runs until the task is done, a limit is hit or a person is needed. A fifth makes it safe to run: a permission boundary deciding which tools exist, what each may touch and which actions wait for approval.

That is an engineering object, not a prompt: it has state, failure modes, a security surface and a bill. This page takes the engineering view. For the process view, which operations to hand over first, see AI automation; if you are still weighing where agentic AI belongs at all, start with AI consulting.

Choosing the right tool

AI agents vs chatbots vs traditional automation

Many requests for an agent are better served by something simpler, and some requests for AI workflow automation need an agent; this is how we sort them.

QuestionTraditional automationChatbotAI agent
Handles unstructured inputNo. Needs clean fields and a known trigger.Yes, but only to answer in text.Yes. Turns emails, PDFs, tickets and messages into actions.
Follows a fixed pathAlways. A new branch is a change request.Loosely. A scripted flow or a free conversation.No. It chooses the sequence at run time, within its tools and rules.
Needs approval stepsRarely. The code path was approved once, in advance.No. With no writes, there is nothing to approve.Yes. Money, customer contact and compliance actions wait for a person until that action type has earned autonomy.
Best forHigh-volume, stable, fully specified processes.FAQ deflection and guided self-service.Multi-step work with messy inputs, several systems and a person who should sign off.

The longer version, with the full side-by-side table, two worked examples and when not to use an agent, is in AI agents vs traditional automation.

Capabilities

What custom AI agent development covers

The reference runtime below shows the layers. These are the three parts a diagram cannot show, and each is where an agent project quietly fails when it is left out.

Tool use and API interactions

Every tool is a typed function with its own schema and credential: search HubSpot contacts, read a Xero invoice, create a Salesforce task. We write the adapters and validate arguments before execution, so a malformed or out-of-scope call fails at the boundary rather than inside your system.

Approval workflows and human-in-the-loop

Approval is a first-class state in the workflow, not a chat message and a hope. The agent packages what it wants to do and why; a person approves, edits or rejects, and the stored decisions show where a threshold can safely move.

Failure handling

Timeouts on every external call, retries with backoff, idempotency keys on writes so a retry never creates a duplicate, fallbacks to a smaller model or a human queue, and a dead-letter queue for runs that still fail, so a failure is visible rather than lost.

Orchestration patterns

Three shapes of agent system

The shape sets cost, latency, testability and how much can go wrong at once. We start with the simplest that fits.

Single agent

One model, one tool set, one control loop: it works through the input and returns a result or asks for approval.

  • The task is bounded and one skill covers it
  • You want the easiest system to evaluate and explain

Workflow with agent steps

A deterministic workflow, usually Step Functions, where most steps are ordinary code and a few are agent calls for the parts that need judgement.

  • The process is mostly known with a few messy steps
  • Auditability and predictable cost are priorities

Multi-agent with a supervisor

A supervisor splits the job, hands each part to a specialist with narrower tools and permissions, then checks and merges the results. Each specialist is evaluated on its own, which keeps the system debuggable.

  • The job spans distinct skills or data domains
  • Specialists need different permission scopes
Reference runtime

A production agent runtime on AWS

The boxes change from project to project; the layers do not, and each needs an answer before an agent goes near a live system.

  • Triggershow a run starts
    Inbound emailWebhooksAPI GatewayEventBridge schedulesQueue messagesA button in an internal tool
  • Control looporchestration and state
    Step Functions state machineLambda stepsECS/Fargate workersSQSPlannerStep and cost budgetsApproval wait states
  • Model and toolsreasoning and action
    Tool registry with typed schemasArgument validationModel router by stepAnthropic, OpenAI and Google via APIBedrock and open-weight modelsStructured outputsConnectors to Xero, HubSpot and Salesforce
  • Memorycontext and knowledge
    Run state in DynamoDBLong-term store in Postgrespgvector retrievalDocument index on S3Permission-aware filtersRetention rules
  • Guardrailspermissions and security
    Least-privilege IAM per toolScoped credentialsAllow-listsInjection screening on untrusted inputExfiltration controlsRate and spend limitsCognito
  • Operationsobservability and delivery
    Trace per runTool-call logReplayCloudWatch alarmsEvaluation harnessTerraform/CDKCI/CD

Our default control loop is a Step Functions state machine with Lambda steps and SQS between them, so a slow tool or a pending approval never holds compute open; heavier workloads move to ECS/Fargate behind the same queues. As in our wider production AI engineering work, the agent is a service with an interface, tests and alarms.

Delivery

How an agent ships

Agent work follows the five stages in how we work. Inside an Implementation block, each agent moves through these five steps.

  1. Map the workflowWalk the job as a person does it today: inputs, systems, decisions, exceptions and who signs off.
  2. Design the tool surface and permissionsDefine each tool, its schema, its credential scope and which actions need approval, before any model is involved.
  3. Build and evaluate on real casesImplement the loop and the tools, then run them against historical cases with known correct outcomes.
  4. Supervised rolloutThe agent proposes, people approve. Every decision is logged, and thresholds move only when the evidence supports it.
  5. Hand it the shift with monitoringApproved action types run without a queue; the rest still wait, with dashboards, alarms and runbooks designed in and documentation handed over.
Use cases

Six agents a business can put to work

Each names the systems it touches and where a person approves, because that is how we scope an agent: autonomy is earned as decisions prove out, never promised where oversight belongs.

01

Sales operations agent

Qualifies new leads against your criteria, drafts the first outreach and books follow-ups.

Systems: HubSpot or Salesforce, email, calendar. Human approves: every outbound message, and stage changes above a value threshold.

02

Research and knowledge agent

Answers staff questions from policies, procedures and past cases with citations, or plans the searches for an open question and returns a brief with contradictions flagged.

Systems: document stores, wiki, web search, a permission-aware index. Human approves: read-only by design; a brief stays a draft until a reader signs it off.

03

Document processing agent

Reads contracts, applications or claims, extracts structured fields, checks them against rules and routes exceptions.

Systems: email and S3 intake, Postgres, your line-of-business system. Human approves: low-confidence extractions and any rejection.

04

Support agent

Reads the ticket and the customer's history, resolves what policy allows and drafts the reply for the rest.

Systems: helpdesk, CRM, account database. Human approves: refunds, credits and account changes always; outbound replies until the decision log earns otherwise.

05

Operations agent

Matches jobs to resources, flags exceptions and prepares the day's schedule for sign-off.

Systems: scheduling, internal database, messaging, calendar. Human approves: schedule changes affecting customers or staff.

06

Finance workflow agent

Checks expense claims and supplier bills against policy and budget, drafts month-end accrual journals and lists anything unusual.

Systems: Xero or MYOB, expense tool, email. Human approves: every journal before it posts, and every payment.

Permissions and security

An agent is a new user in your systems

An agent that can write to your CRM, send an email or create an invoice is a new principal in your environment, and we treat it as one. The threats specific to agents are prompt injection carried in what it reads, data exfiltration through wherever its output can go, and tool abuse; each has a named control, not a hope that the model behaves. We recommend a security review before any live credential is issued.

Designed in by default

  • Against tool abuse: scoped, least-privilege credentials per tool, never a shared admin key, and no general-purpose tool for arbitrary API calls, so a capability outside the registry has no route to it, however the agent is prompted
  • Against injection: instruction and data kept separate, with untrusted content screened before the model reads it
  • Against exfiltration: allow-lists on recipients, records, amounts and endpoints, so output can only go where the job needs it to
Why ComplxAI

Why build agents with ComplxAI

Agentic and automated systems we have shipped

Modedu: agentic workflows across operations and an agentic-first custom whiteboard. Luxpip: agentic solutions for internal ops. Billy: agentic workflows through core operations. DisabilityAssessments, automation rather than agents: matching and alerting, invoice processing and reconciliations, with 90% of operations automated.

The runtime, not just the prompt

Tool adapters, state machines, memory stores, evaluation harnesses and alarms are built by the senior team that designed them, working remote-first across Australia and onsite where it helps. When the agent needs an application around it, our custom AI development work covers the rest.

Yours from the first commit

The tool registry, prompts, state machine definitions and Terraform are committed to your repositories from day one of any paid stage and deployed into your own AWS account, or hosting we set up for you. Handover and documentation are included; take the runtime in-house at any boundary.

Priced per stage

Two free conversations, then Discovery, Strategy and 12-week Implementation blocks, each priced in writing before it starts. Exit at any boundary. No mandatory retainers.

Questions

Questions buyers ask about AI agents

Bring the specific job and the systems it lives in.

Book a free intro call
How do you stop an agent doing something wrong?

By limiting what it can do before worrying about what it might decide: scoped credentials and allow-lists per tool, approval gates on consequential actions, step, time and spend budgets on every run, and a replayable trace. It runs supervised first, and autonomy widens one action type at a time.

What does it cost to build an AI agent?

We price the stages, not the agent: Discovery from a fixed fee, Strategy priced upfront and 12-week Implementation blocks, each in writing before it starts, and the AI project cost calculator gives a rough range.

Which models do you use for agents?

Whichever passes the evaluation for each step at the lowest cost: OpenAI, Anthropic and Google APIs, Amazon Bedrock, or open-weight models where data handling favours them. A router lets a small model classify and extract while a larger one plans, and the tool and memory layers do not change when the model does.

Can an agent work with Xero, HubSpot or Salesforce?

Yes. Those three, with MYOB, email, calendars, document stores and internal databases, are the integrations we are asked for most often. Each becomes a typed tool with its own credential and allow-list; our AI integration services page covers that layer.

How long until the first agent is running?

Discovery takes 2–4 weeks, Strategy another 2–4, and Implementation runs in 12-week blocks with working software every couple of weeks. We recommend scoping the first block around one agent in supervised operation. Two things gate the timeline: access to the systems the agent will touch and agreement on the approval rules, so both start in Discovery.

Start with a conversation

Tell us the job you want handed to an agent

Bring the process, the systems it touches and the decision you would never let software make alone. Two free conversations first; every stage after that is priced in writing before it begins.