Governance Becomes the Product
For most of the past two years, the enterprise conversation about AI agents has been about capability: what they can do, how autonomous they can be, how quickly a team can stand one up. Thoughtworks is arguing that the conversation has to change, and its new platform, Agent/works, is the argument made concrete. The pitch is that the hard problem is no longer building agents, it is running them safely at scale, and that the industry has been treating governance as an afterthought precisely when it can least afford to.
The premise is straightforward and uncomfortable. When anyone in an organization can spin up an agent that reads data, calls tools, and takes actions, the old model of reviewing software before it ships breaks down. Agents are dynamic, they make decisions at runtime, and a static approval gate cannot see what they will actually do once they are live. Agent/works is Thoughtworks' attempt to move oversight from the planning stage into the moment of execution, where the risk actually lives.
What Agent/works Actually Does
At a functional level, Agent/works provides a single control plane and a governed runtime for managing AI agents across any cloud environment. That any cloud framing is deliberate. Enterprises are not going to standardize on one model provider or one hyperscaler, and a governance layer that only works inside a single vendor's walls would be dead on arrival. The platform is designed to sit above that heterogeneity and give teams one place to see and control the agents scattered across it.
The capabilities Thoughtworks emphasizes map directly to the failure modes organizations are already hitting. The platform offers provable compliance before execution, capability based permissions that constrain what an agent is allowed to touch, a governed runtime that enforces those constraints as the agent works, a composable design, and a centralized registry that gives teams fleet wide visibility. Taken together, that is an answer to the two questions most enterprises cannot currently answer: which agents are running, and what are they permitted to do.
The Numbers Behind the Urgency
The timing is driven by a curve that is bending fast. Gartner projects that 40 percent of enterprise applications will have embedded agents by the end of the year, up from less than 5 percent in 2025. That is not gradual adoption, it is a step change compressed into a handful of quarters, and it means governance frameworks designed for occasional pilots are about to be overwhelmed by production reality.
The gap that opens between capability and control is where the risk concentrates. Every agent added to an enterprise application is a new actor that can read data, invoke services, and consume resources, often with permissions that were granted quickly and never revisited. Multiply that across thousands of applications and the result is agent sprawl: a population of autonomous processes that no single team fully understands. Thoughtworks is betting that the enterprises feeling that pain first will pay for a way to regain visibility before something goes wrong.
Governance as Runtime, Not Checklist
The philosophical core of Agent/works is a rejection of governance as paperwork. Shayan Mohanty, Thoughtworks' chief data and AI officer, put the case directly. When anyone can generate software with a text prompt, he argued, AI governance is not a checklist you bolt on after the fact. It is a foundational requirement for operating autonomous systems at scale, and the implication is that oversight has to be built into the system rather than layered on top of it after deployment.
Mohanty also pushed back on the assumption that governance and speed are in tension. Governance done this way, he said, stops being a brake and becomes the engine that lets teams move fast, safely and at scale. That is the reframing Thoughtworks is selling: guardrails that are embedded in the runtime let engineers ship with confidence rather than waiting on manual review. Whether the market accepts that trade will depend on whether the controls genuinely accelerate delivery or simply add a new layer of process wearing a runtime costume.
The Cost Meter No One Budgeted For
One of the less discussed problems Agent/works targets is economic, not just architectural. Every AI powered workflow now carries an operating cost, and Mohanty framed the challenge as no longer just how to build agents, but how to govern the resources they consume. Agents that call models and tools repeatedly can run up bills that surprise finance teams, and without fleet level visibility those costs are nearly impossible to attribute or control.
David Nasi, a director of product management at Databricks, pointed to where the pattern is heading. The organizations moving fastest with agents, he observed, are extending the same governance models already used for enterprise data across agent workflows themselves. That analogy is instructive. Enterprises spent the last decade building data governance because ungoverned data became a liability, and agents are following the same arc. The companies that treat agent governance as infrastructure, rather than as an obstacle, are the ones setting the pace.
Why This Lands Now
Agent/works arrives at a moment when the enterprise mood around AI has shifted from experimentation to accountability. The pilots have run, the demos have impressed, and now boards and regulators are asking harder questions about what these systems do when no human is watching. A platform that promises to answer who is running which agent and what it is allowed to do is aimed squarely at that anxiety, and it reflects a broader market signal that the money in 2026 is moving toward integration and control rather than more proofs of concept.
The open question is whether governance can be productized this cleanly, or whether every enterprise will insist on shaping its own controls to fit its own risk posture. Thoughtworks is wagering that a common control plane is valuable enough to overcome that instinct, much as shared platforms eventually won out in data and infrastructure. We would watch adoption among the organizations already drowning in agent sprawl, because they are the ones with the clearest reason to buy rather than build. If Agent/works finds traction there, it will validate the thesis that in the agentic era, governance is not overhead, it is the platform.
