Crossing the agentic chasm

The emerging divide

Copilot did something important. It put useful AI inside applications people already knew how to use. Writing, summarizing, finding information, preparing meetings, working through documents: none of that is trivial, and none of it disappears because the market has started talking about agents.

But the people pushing hardest are beginning to work differently. They are moving into what I call harness-first applications: Claude Code, Codex, Cursor, and similar environments built around a persistent workspace. Context, instructions, tools, files, tests, and prior decisions stay close to the work. A user can delegate something substantial, inspect the result, and reuse what worked. The attraction is not a better chat window. It is the ability to build operating capacity around a continuing body of work.

The category has fuzzy edges, and GitHub Copilot increasingly works this way in software development. The difference is where the experience starts. A copilot starts beside the user inside an application. A harness starts with the work and asks which models, tools, and systems it needs.

This has exposed a tension that the market has not resolved. Microsoft has distribution: Copilot sits inside the applications, identity boundaries, and data estate that companies already use. Harness-first applications give practitioners more freedom to assemble context and execution around the problem in front of them. The first path can spread AI broadly without changing how work is organized. The second can change the work before the company knows how to operate it.

The chasm is between what a skilled user can make AI do and what an enterprise can responsibly operate.

The two curves are pulling apart

Tool adoption moves quickly because the feedback loop is personal. Someone tries an AI tool, finds a useful task, changes how they work, and keeps going. Copilot has made that loop broadly available. Harness-first applications give advanced users more room to discover workflows that a central program would not have designed in advance.

Enterprise transformation moves differently. An impressive result has to become a repeatable workflow. The company needs to know what can be delegated, what counts as done, where judgment remains human, who owns a failure, and which system records the result. Permissions, support, measurement, and change control all enter the picture.

The two curves reward different things. Tool adoption rewards low friction, personal configuration, and room to improvise. Enterprise transformation rewards repeatability, stable interfaces, and clear ownership. A company cannot simply copy its most capable users, but it also cannot standardize the future before they discover what is worth scaling.

The market is further along the first curve than the second. AI use has spread faster than operating models have changed. Many organizations can point to impressive individual examples. Far fewer can show that a workflow that matters has been redesigned, governed, measured, and operated repeatedly across teams.

This gap creates two easy mistakes. Microsoft can mistake licenses and frequent use for transformation. A harness vendor can mistake an expert user's breakthrough for something an enterprise can run. Crossing the chasm means preserving the freedom to discover better work, then making the useful discoveries dependable. That changes both how value is measured and what the company must trust.

The unit of value changes

When an agent accepts a goal and acts through tools, the unit of value changes. Tokens, seats, prompts, and generated documents tell us what was consumed. They do not tell us what the organization received.

AI consumption appears in a technology budget, while the value may show up elsewhere: engineering finishes sooner, a service is avoided, risk falls, or a customer gets an answer faster. Old cost centers make the expense visible and the value hard to see.

I keep coming back to cost per accepted outcome. What did it cost to assemble the context, run the work, check it, handle failure, and get an accountable person or system to accept the result? What human effort, delay, service expense, or risk did it remove? A cheap answer that needs hours of reconstruction is expensive. A high-consumption run that clears a costly bottleneck may be a bargain.

This is where agents begin to look like operating capacity rather than another software tool. Harness-first applications keep enough context to avoid starting over, make instructions and tools reusable, and retain the history of the work. Measuring that only as an IT meter leads to precise decisions about the wrong thing.

This changes what any AI platform has to prove. More seats and more usage are not enough. A portfolio has to help customers produce accepted outcomes and turn the best workflows into operating capacity. But an accepted outcome only matters at scale if the company can trust how it was produced.

Delegation crosses a trust boundary

Harnesses gain power by carrying context across tools. If I can reach a Power BI report or operate an application, a harness may be able to bring that information or action into the workflow. This speeds up experimentation, but it also crosses a trust boundary. Human access does not automatically permit extraction, storage, combination with other sources, or delegation to another model or agent.

The ability to click through an application also does not make the automation supported or reliable. APIs remain the safer foundation. Where UI automation is necessary, the agent needs a clear scope, a way to verify the result, and a safe route to stop or recover.

For a global company, the delegated actor needs an identity, policy that follows the data and action, and data that is reliable, current, allowed for the task, and traceable to its source. The company also needs consistent business definitions, logs, cost and permission controls, and an owner when the workflow fails.

Those controls turn a clever demonstration into a service. They also define the gap each side of the market is trying to close.

A race in both directions

The large platforms have identity, data, compliance, distribution, procurement, and established systems of record. The harness-first vendors have practitioner preference, continuity, configurability, and a shorter path from experimentation to reusable work.

Both positions are incomplete, but for different reasons. Central control can distribute an assistant widely without changing the underlying workflow. Unbounded flexibility can leave the company with stale context, scattered permissions, brittle automation, and critical knowledge trapped with the person who built it.

So the race runs both ways. Bottom-up vendors need to become governable before the incumbents become genuinely agentic. The incumbents need to create an experience people choose before the new harnesses gain enough enterprise control to become safe defaults.

Public evidence does not yet show a market-wide move away from any one assistant. But advanced-user demand is the leading indicator to watch: where the most capable users go first is often where the enterprise follows once the controls catch up.

The early majority will expect both. It will want the power to change the work and something the company can rely on. The work has to repeat. The outcome has to be verifiable. The economics have to survive real failure rates and human supervision.

Start with one workflow

The path to durable value begins with a workflow, not a product diagram. Find the people already getting unusual results. Describe the job, data, tools, judgment points, completion test, and final system of record. Then define the identity, context, semantics, and governance needed to reproduce it safely.

Pilots should be built around acceptance tests, not demonstrations. Did the agent reach the required state at an acceptable cost and quality? Can someone inspect the evidence without replaying the whole session? What happens when a tool fails or the data conflicts?

Production needs supported integrations, permission boundaries, exception handling, rollback, telemetry, evaluation, and an owner. People must learn to set outcomes, review evidence, handle exceptions, and improve the workflow. AgentOps then manages drift, cost, incidents, and lifecycle.

License activation is only an input. The next step is an accepted outcome. Transformation begins when a team repeats the workflow often enough to reorganize work around it, then carries the pattern into the next part of the company.

Crossing the chasm

The chasm will not be crossed by distributing more assistants or by giving agents more freedom. It is crossed when a workflow discovered by a skilled user can be repeated across teams, accepted at a cost that makes sense, and operated within the company's trust boundary.

This is where the platform opportunity sits. Harness-first applications may lead the discovery of new ways of working. The platforms that endure will be the ones that connect those breakthroughs to identity, governed data, systems of record, security, runtime, and the work surfaces used across the enterprise. The assistant must compete for the user experience, but the platform can remain central to how the work is operated.

The job is to make that progression real with customers now: prove value in the current workflow, define the accepted outcome, add the controls needed to repeat it, and expand from there. No vendor will cross the agentic chasm through distribution or practitioner preference alone. The advantage belongs to whoever can connect the two.