Software that is given a goal rather than a script. It plans the steps, uses tools such as search, databases and internal systems, checks its own progress, and reports back when the goal is met or blocked.

How is that different from automation we already have?

Traditional automation follows a fixed path and fails when reality differs. An agent re-plans when a step fails, which is what makes it useful for messy, exception-heavy work.

What stops an agent from doing something harmful?

Every tool an agent can use is explicitly permitted, scoped and logged. High-impact actions require human approval, and AEGIS supervises the agent's plans as well as its outputs.

Can several agents work together?

Yes. Work is split across specialist agents — one researches, one drafts, one verifies — coordinated by a supervisor that holds the overall goal and resolves conflicts.

What happens when an agent gets stuck?

It stops, records what it tried and hands the task to a person with full context, rather than looping indefinitely or producing a confident but wrong result.

All modules

M04 · In Build

HELIX

Autonomous agent orchestration layer

Investment

$2,100,000

Share of programme

14.7%

Timeline

Q1 2026 — Q4 2027

Delivery team

16 engineers

35% complete

Autonomous agent orchestration graph visualisation for the HELIX module — M04 · HELIX — Autonomous agent orchestration layer

In plain English

Turns answers into completed work. Agents plan a task, use real tools and systems to carry it out, and stop for human approval where the stakes are high.

What the software does

Breaks a request such as 'reconcile this month's supplier invoices' into an ordered list of steps.
Uses real systems — databases, internal APIs, spreadsheets, ticketing tools — to actually perform those steps.
Checks its own output, retries what failed and escalates what it cannot resolve.
Pauses for a named human to approve anything irreversible or above a value threshold.

How it works, step by step

01PlanThe goal is decomposed into steps, each with a defined success condition rather than a vague instruction.
02Act with toolsThe agent calls the systems it has been explicitly granted access to, and every call is logged.
03Verify and hand backResults are checked against the success conditions; anything uncertain or high-impact goes to a human before it is committed.

A simple analogy

Less a chatbot, more a capable junior colleague who works the queue, keeps notes, and knocks on your door before spending real money.

Why it matters

Customers pay for completed outcomes, not for text. Agents are where the platform converts model capability into measurable labour savings.

How HELIX works

Inside the module

A goal goes in; completed multi-step work with a replayable trace comes out.

Input

Goal
What the user wants achieved, stated in plain language
Tools
Permitted systems, databases, APIs and connectors
Rules
Business policy, approval thresholds and limits

M04 pipeline · select a stage

1/4

The goal is broken into an ordered set of steps

Instead of following a fixed script, the agent works out a sequence of actions that should reach the goal, and identifies which tools each step needs.

Task decompositionTool selectionDurable plan state

Output

Completed work
The task actually done, not just described
Run trace
Every step, tool call and decision, replayable
Escalations
Blocked tasks handed to a person with full context

Stage by stage, in detail

Declare

Agents are defined with goals, permitted tools, spend budgets and escalation rules — no bespoke code per workflow.

Plan

The planner decomposes long-horizon goals into checkpointed steps with explicit success criteria.

Execute durably

A workflow engine persists state, so multi-day runs survive restarts and partial failures.

Gate

Human-in-the-loop approvals interrupt any step that crosses a risk or cost threshold.

Replay

Every input, tool call, reasoning summary, cost and approval is recorded as a replayable trace.

Registered tools

Internal and third-party integrations in the tool registry

Multi-day

Workflow horizon

Alpha engine runs without state loss across restarts

100%

Traced steps

Every agent action is recorded and replayable

Durable workflow engineTyped tool registryPermissioned memory serviceBudget governorReplayable trace storeApproval and escalation framework

Questions answered

HELIX FAQ — how the AI works, in plain terms

Common investor questions about what this module does, how it does it, and why it is funded as part of the programme.

Scope

HELIX turns the model from an answering system into a working system. Agents are defined declaratively with goals, permitted tools, budgets and escalation rules, then executed on a durable workflow engine that survives restarts and partial failures.

Every step is recorded as a replayable trace: inputs, tool calls, intermediate reasoning summaries, cost, and the human approvals that gated it. That trace is what makes autonomy auditable enough for regulated buyers.

A shared memory service gives agents persistent, permissioned recall across sessions and across teams.

Contracted deliverables

Durable agent execution engine with replay
Tool registry with typed schemas and permission scopes
Persistent, permissioned agent memory service
Human-in-the-loop approval and escalation framework

Achieved to date

Execution engine alpha running multi-day workflows without state loss
Tool registry covering 60 internal and third-party integrations

Currently in production

Multi-agent negotiation and hand-off protocol
Cost and budget governor per workflow
Failure-recovery planner with automatic retry strategies

Where it is used

Back-office automation

End-to-end processes that span several systems of record with an auditable record of every action.

Research and diligence

Long-horizon investigation tasks with citation trails and controlled spend.

Regulated autonomy

Sectors that can only deploy autonomy when every decision is reconstructable after the fact.

Platform dependencies

Reasoning from NEXUS
Perception via ORACLE
Policy enforcement from AEGIS
Exposed to users through ATLAS

Key risks and mitigations

Unbounded agent cost

Per-workflow budget governor halts execution at a hard ceiling.

Failure cascades in long runs

Checkpointed recovery planner with automatic, bounded retry strategies.