Atomicwork CEO: A genuine architectural leap - building agents with “job roles”

A different animal

Why does this question matter? Because, says Atomicwork. CEO’s name is Vijay Rayapati, production is a different animal entirely i.e. putting an agent into live production means building software that behaves like a permanent member of the workforce: making decisions, touching multiple systems, collaborating with human colleagues and running continuously inside an organisation’s environment without falling over.

Rayapati draws his thoughts on this subject at the head of Atomicwork, a company for its work modernising IT service management with AI-powered software that automates employee support and streamlines enterprise workflows directly inside tools like Slack and Microsoft Teams.

Infrastructure first, then agents

The key now (or so it appears) is all about building the infrastructure that lets agents operate safely, reliably, and at genuine enterprise scale.

“This matters because production AI behaves rather more like an employee than it does like conventional software. Traditional applications execute a predefined set of instructions and nothing more. AI agents, by contrast, perform work, access multiple systems, talk to people and run long, adaptive workflows that shift as conditions change,” said Rayapati.

Rayapati: Reliable enterprise AI now depends on orchestration, enterprise context, identity, policy enforcement, accountability, evaluation, telemetry, workflow execution and interoperability across business systems.

He advises that these realities exist because once an agent moves past answering questions and starts actually doing the job, it needs the same “operational scaffolding” a human employee would i.e. identity, permissions, defined responsibilities, approved tools, organisational context, and clearly drawn boundaries around what it is and isn’t allowed to do.

Beyond prompts

Atomicwork’s approach is to treat its AI coworkers as role-based entities (complete with shiny new identities) as well as job roles, skills, lifecycle management analysis, budgets and performance metrics. These are fully fledged agents, rather than mere bundles of prompts.

Building the operating system around the model has become the primary engineering challenge, not the model itself.

“Reliable enterprise AI now depends on orchestration, enterprise context, identity, policy enforcement, accountability, evaluation, telemetry, workflow execution and interoperability across business systems – the platform wrapped around the model is increasingly what decides whether that AI succeeds or quietly fails in production,” explained Rayapati.

He says that Atomicwork’s architecture reflects this with multiple models stitched together with enterprise context, automation engines, MCP support, evaluation frameworks, real-time telemetry, and hierarchy and identity controls.

What developers should think

Rayapati suggests that software application developers should stop thinking in terms of tasks and start thinking in terms of job roles.

“Most AI deployments begin small – resetting passwords, routing tickets, summarising reports, which are useful, but not transformative,” said Rayapati. “The next generation of AI systems will instead be designed around whole job roles. A network operations engineer, a cloud operations manager, an access manager – these are defined by goals, responsibilities, tools, permissions, escalation paths and measurable outcomes, not by a single tidy workflow.”

Designing agents for job roles means connecting multiple workflows under clearly governed operational boundaries; it’s a genuine architectural leap from workflow automation to autonomous execution.

“All of which means managing AI is going to start looking a great deal like managing infrastructure. Deploy dozens, or hundreds, of AI workers and the operational questions become engineering questions fast: which agents can touch sensitive systems, which version is live, which tools each one is permitted to use, how performance gets measured, when an agent should be updated or retired, and how budgets are enforced,” said Rayapati.

He thinks that these are the same disciplines engineers already apply to infrastructure, identities, applications and cloud resources. Further, he advises that “the real long-term challenge” was never deploying one more agent – it’s building the governance, observability, lifecycle management and operational control needed to let large fleets of AI workers operate safely, together.

Manage AI workforces, not just models

The upshot just might be that the next generation of developer platforms will manage AI workforces, not just AI models.

“The first generation of enterprise AI proved intelligent software was possible,” asserted Rayapati. “The next has to prove it can be dependable. Success will hinge less on who has the latest frontier model and more on who has built the platform to manage, observe, secure and hold that AI accountable across its lifecycle. Developers will spend increasingly less time bolting on another model integration, and increasingly more time building orchestration layers, governance frameworks, identity systems, evaluation pipelines and execution environments.”

In closing, Rayapati tells us that the companies that define this next era won’t necessarily be the ones with the smartest models (and you know what’s coming next by this point), instead, they’ll be the ones that build the infrastructure allowing AI to operate as a genuinely trusted member of the workforce.