Client storyAI platform · Workflow engineering
Auxi Labs

Workflows that can think beyond the graph.

Auxi Labs rebuilt its execution engine so teams could create workflows in plain language and the system could adapt as work unfolds.

See what changed

Inside the workflow

Auxi Labs

  1. 01Bring context into the run
  2. 02Make execution durable
  3. 03Describe the workflow

The change

Workflows in plain language

Client

Auxi Labs

Use case

Workflow execution engine

Stage

Platform evolution

01 / The challenge

The work before the agent

Auxi Labs builds a white-label AI agent platform for enterprise workflow automation. Its original engine used directed acyclic graphs (DAGs) to define each workflow in advance.

As workflows grew more complex, that fixed path became a constraint. Different user contexts required separate definitions. A failed step could halt the run or trigger a blind retry, and each workflow change still needed engineering time.

02 / The approach

What we built

We replaced static orchestration with an agentic architecture that uses context to decide what happens next, while keeping execution durable and reviewable.

01

Bring context into the run

A Context Fabric knowledge graph holds what the platform knows about users and teams. Workflows can prefill known details and adapt to the situation.

02

Make execution durable

Restate manages workflow state and recovery. Timeouts can be retried, validation issues routed back to a user, and a run paused through a service outage.

03

Describe the workflow

MCP servers translate a natural language description into a structured, runnable workflow. Implementation teams can change the flow without defining another graph in code.

04

Adapt at each step

The agent uses the available context and the result of each action to choose the next step, handle exceptions and improve execution patterns over time.

03 / The result

What changed

For end users

Known details can be filled in, and workflows can respond to the situation or recover from a failed step.

For implementation teams

Teams can create and iterate on workflows in plain language instead of requesting a new graph for every variation.

For engineering

Fewer small workflow changes and failed runs need manual intervention.

The execution engine can now work with context that was previously trapped outside a fixed workflow definition. This changes who can create workflows and how the platform deals with exceptions.

The architecture combines adaptive decisions with durable state, so flexibility does not depend on losing track of a run when something goes wrong.

Under the hood

The implementation

A closer look at the systems and controls behind this story.

  • Agentic execution engine
  • Context Fabric knowledge graph
  • Restate durable execution
  • MCP servers
  • Context-aware routing

Keep exploring

More work in production

All case studies →
Confidential operations clientRecurring work automatedInput checks, approved actions and a completion report for review.Read story →Australian govtech clientSelf-service data accessCouncil staff can explore data without requesting a custom report.Read story →

Your workflow next

What would your workflow engine need to handle?

Show us the step that keeps failing. We will trace the handoff and work through what the engine needs to do differently.

Discuss a workflow
Discuss a Workflow