From agentic workflows to an agentic workforce

Last week, OpenAI launched Frontier. If you skimmed the announcement, you’d be forgiven for thinking it was another enterprise AI platform. Another press release, another set of logos, another promise about productivity.

But the language in that announcement stopped me cold.

“AI coworkers.” “Onboarding.” “Identity.” “Shared context.” “Institutional memory.” “Permissions and boundaries.”

That’s not automation language but proper employment language.

For most of 2025, the industry conversation centred on agentic workflows. Agents that could plan, reason, and execute multi-step tasks autonomously. Every major vendor built a platform around this idea: Salesforce launched Agentforce, Microsoft expanded Copilot Studio, Google rolled out Agentspace, AWS shipped Bedrock AgentCore. All useful. All genuinely impressive. But all framed around the same basic metaphor: process automation. The unit of deployment was a workflow. The organisational model was bolt-on. You had your existing processes, and you strapped agents onto them.

Frontier is different. When you give agents their own IAM identities, persistent memory, onboarding processes, performance evaluation, and explicit roles alongside human employees, you’ve crossed a line. You’re no longer automating a process. You’re staffing a position.

And I don’t think we can uncross it.

The scaffolding matters more than the capabilities

I wrote previously about AI as a coworker from a capabilities perspective (back when Gemini 3 was released). What could these systems actually do? How good were they at reasoning, at writing, at handling complex tasks? Those questions still matter. But something has shifted since then, and it’s not about what agents can do. Now, an infrastructure is being built around them to make them permanent members of an organisation.

Think about what Frontier actually includes.

A semantic layer that connects data warehouses, CRM systems, ticketing tools, and internal apps so agents understand how information flows. Agent execution environments with persistent memory from past interactions. Built-in performance management where human managers evaluate AI coworkers the same way they’d evaluate a new hire. Enterprise identity and access management that applies across both human employees and AI coworkers. Even forward-deployed engineers from OpenAI who embed with customer teams to help with “onboarding.”

That last bit is fascinating. OpenAI is sending people to help you onboard your AI coworkers. Let that sink in for a moment.

OpenAI Frontier's architecture diagram tells the story. The bottom three layers (Business Context, Agent Execution, Evaluation and Optimisation) are the scaffolding that turns agents from isolated automations into managed workforce members. Notice "Your Agents" sitting alongside "OpenAI Agents" and "Third-Party Agents" — the platform is designed to manage a blended agent workforce, not just OpenAI's own models. The "Enterprise security and governance" wrapper on the right is doing the work that HR and compliance do for human employees. Source: OpenAI.The point is this: capabilities without scaffolding produce interesting demos. Scaffolding is what produces organisational change. And the scaffolding being built right now is explicitly designed to treat agents as workforce members, not workflow components.

McKinsey’s Jorge Amar put it plainly: “Some pioneering companies in this space are expressing their org charts not only in number of FTEs but also in number of agents being deployed in every part of the organization.” He went further, describing companies pursuing “zero-FTE departments” where entire functions are performed by agents, with humans monitoring and controlling on the side.

Jensen Huang’s version was a bit punchier: “IT will be the HR of AI agents of the future.”

I keep thinking about this from a content and marketing perspective, because that’s where I live. For the past year, most of us in content operations have been building agentic workflows. An agent that drafts product descriptions. Another that checks copy against brand guidelines. A QA pipeline that flags errors. These are useful. I’ve built several of them myself. But they’re still isolated automations bolted onto existing processes. What changes when you start thinking of these agents as team members with defined roles, persistent context about your brand voice and product catalogue, and performance metrics that sit alongside your human writers’ KPIs?

Everything changes.

The org design changes.

The management model changes.

The way you think about hiring changes.

What I’m seeing from inside a large organisation

I’ve had the chance, from my position managing AI automations within a large retail group, to test and assess how the major vendors are actually thinking about this. And the approaches are more different than the marketing would suggest.

Gemini Enterprise, from my direct experience, is taking an approach that rhymes with what OpenAI is doing with Frontier. Agentspace isn’t just an agent builder, but rather an environment designed for deploying enterprise-grade agents across Gmail, Docs, Sheets, Chrome, and third-party apps, with a no-code Agent Designer and a centralised Agent Gallery. The A2A protocol (Agent2Agent) is an open standard for cross-platform agent communication. The thinking is systemic. Google wants agents embedded across the organisation’s entire digital surface, communicating with each other, governed centrally. It feels like workforce planning, not workflow design.

Microsoft is taking a similar direction conceptually with Copilot Studio and Agent 365, but from a user experience perspective, they feel a step behind. The capabilities are there. Multi-agent orchestration, multi-model support, natural-language agent creation. Agent 365 introduces a centralised governance layer specifically designed to address agent sprawl. Their own analysis includes the prediction that “autonomous agents will soon outnumber human users inside enterprise systems.” They clearly see where this is going. But the experience of actually using these tools day-to-day doesn’t yet match the ambition of the vision. The plumbing is solid; the interface needs work.

Then there’s Anthropic, and I think they’re doing something genuinely different. Claude Cowork launched in January, and rather than building an agent management framework for IT teams to deploy fleets of workers, Anthropic handed an autonomous agent directly to end users. It lives on your desktop. It manages files, drafts documents, and handles tasks. Built on the same Claude Agent SDK that powers Claude Code (which developers already love), but packaged for non-technical people.

Claude Cowork offers most of the capabilities that were initially accessible to developers via Claude Code, but wrapped in a user-friendly UI. Claude can access and edit your files and do real work on your computer, unlocking use cases such as the ones depicted in this screenshot. Source: Anthropic.Something is interesting in that divergence. OpenAI and Google are building top-down: give the organisation a platform to deploy and manage an AI workforce. Anthropic is building bottom-up: give individual workers an AI colleague so capable that the organisation has to adapt around it. Both approaches end up in the same place eventually, but the routes are very different, and the implications for adoption are too.

Anthropic’s MCP (Model Context Protocol) is already hitting 100 million monthly downloads and becoming an industry standard for how agents connect to tools and data. That’s the kind of infrastructure play that could make Cowork’s bottom-up approach sticky in ways that are hard to unwind.

From a content and marketing leader’s perspective, I find the Anthropic approach the most immediately useful. When I’m trying to get a team of eight writers to change how they work, handing each of them a capable AI colleague is a much easier sell than asking IT to deploy an enterprise agent fleet. The adoption path is more human. But the governance path is messier. And that’s the trade-off every organisation is going to have to navigate.

The competitive landscape is a land grab

Everyone is racing to own what I’d call the orchestration layer. The platform that sits between an organisation’s data, its human workforce, and its AI agents. Whoever controls that layer controls the enterprise relationship.

The numbers explain the urgency. The agentic AI market is projected to grow from $7.29 billion in 2025 to $139.19 billion by 2034, a CAGR of 40.5%.

And the financial markets are responding accordingly. When Claude Cowork launched in January, legacy SaaS stocks dropped 14% in a single week. Analysts started calling it the “SaaSpocalypse.” Whether that label sticks or not, it signals something real: the market believes AI agent platforms will displace conventional enterprise software, not just augment it.

The vendor lock-in tension is real, too. Tatyana Mamut, CEO of agent observability company Wayfound, captured the mood: “Nobody is signing multi-year contracts anymore because if something great comes out next month, I need to be able to pivot, and I can’t be locked in.” OpenAI claims Frontier is open and can manage agents built outside their ecosystem, but hasn’t clarified whether it’ll support third-party models. Multi-model flexibility is becoming table stakes. Both Salesforce and Microsoft already offer it. If OpenAI doesn’t follow, that’s a vulnerability.

The governance problem nobody’s ready for

Here’s where my optimism gets complicated.

Most organisations don’t have a central view of their AI agent inventory. Teams build duplicate agents for the same tasks without knowing that others exist. There’s no unified visibility, no consistent controls, no governance framework. This is being called agent sprawl, and it’s already happening.

And then there’s the identity problem. Traditional identity and access management were built for human users with stable, long-lived identities. AI agents are ephemeral. They spin up and down in seconds. They might need access to different systems depending on the task. They don’t have a physical badge or a manager who can vouch for them.

The Cloud Security Alliance has published a purpose-built framework for agentic AI identity management using decentralised identifiers, verifiable credentials, and zero trust principles. Each agent gets a dedicated identity tied to a verified human or organisational owner. Just-in-time, task-specific credentials instead of permanent passwords. Automated provisioning and de-provisioning. Sponsors or custodians who review and certify agent access.

It reads like an HR policy document. Because that’s essentially what it is.

Frontier addresses this directly, with enterprise IAM that applies across both human employees and AI coworkers.

I think that the companies that will do well are the ones that will create dedicated centres of excellence. Not just for building agents, but for governing them. Discovery, registration, access controls, performance dashboards, and retirement processes.

In other words, this is workforce management. The vocabulary choice is deliberate.

The human question is genuinely unsettled

I’ll be honest: the workforce impact data pulls in both directions, and I don’t think anyone has a clean answer.

The World Economic Forum projects that by 2030, job disruption will affect 22% of all jobs, with 170 million new roles created and 92 million displaced. A net gain of 78 million positions globally. That sounds reassuring until you zoom in.

Dario Amodei, Anthropic’s own CEO, suggested in May 2025 that AI could eliminate half of entry-level white-collar jobs and spike unemployment by up to 20% over five years. Ford’s Jim Farley suggested a halving of demand for white-collar work. Klarna’s CEO stated that their AI chatbot does the work of 700 customer service agents.

MIT economist David Autor offers the counter-view: “The industrialized world is awash in jobs, and it’s going to stay that way.” An Economic Innovation Group analysis showed that occupations exposed to AI don’t exhibit elevated rates of unemployment.

Both things can be true simultaneously. Aggregate job creation can mask brutal displacement in specific roles, industries, and geographies. As someone who manages a team of writers and builds AI automations for content creation, I’m acutely aware of this tension. I see the productivity gains. I also see what gets lost.

New roles are also emerging.

Agent manager. AI compliance officer. Human-AI collaboration designer. The McKinsey analysis describes a world where “workflows must be designed for people managing agents rather than people doing tasks.” That’s a genuine restructuring of work, and it has real implications for hiring, training, career development, and organisational design.

Where this actually leaves us

OpenAI’s framing is that “the primary constraints for organisations are no longer model performance or tooling, but rather organisational readiness and implementation.”

I think that’s right, and I think it applies equally to the content and marketing functions I work in.

The models are good enough. The platforms are arriving. The scaffolding for treating agents as workforce members is being built in real time by OpenAI, Google, Microsoft, Salesforce, AWS, and Anthropic.

The shift from agentic workflows to an agentic workforce is real. I can see it in the platforms being built, the language being used, the organisational structures being proposed, and the governance frameworks being drafted. I can feel it in my own work, where the question is increasingly not “can an agent do this?” but “how do we integrate agents into the team in a way that actually works?”

The answer to that question is going to be different for every organisation.

But the question itself has changed. We’re not asking whether to automate a process anymore. We’re asking how to staff a blended workforce of humans and AI. And the companies that figure out the governance, the identity management, the change management, and the basic human decency required to navigate this transition are the ones that will come out the other side intact.

What I find most interesting, and slightly uncomfortable, is how quickly this shift happened.

Twelve months ago, we were talking about agentic workflows as the cutting edge. Now we’re talking about agent onboarding, agent identity, and agent performance reviews. The vocabulary moved from DevOps to HR in less than a year.

That speed should make us both excited and cautious.

Excited because the scaffolding being built is genuinely sophisticated and could unlock real organisational transformation. Cautious because we’ve been here before with enterprise technology, and the gap between platform capability and organisational readiness has historically been where ambition goes to die.

The rest will have a lot of agents and very little to show for it.

If you’re working through how agentic AI fits into your content operations or broader organisational strategy, reach out through cowritten.ai.