Organizations Must Invest in AI Oversight

CHICAGO and LONDON, Nov. 4, 2026 /CourierPR/ -- Thoughtworks, a global technology consultancy that integrates design, engineering, and AI to drive digital innovation, today released volume 35 of its Technology Radar, a biannual report informed by the organization's experiences with clients. The release highlights the growing importance of safeguards needed to run AI reliably in production, as the adoption of AI agents scales.

The 11 practices and technologies placed in the Technology Radar's Adopt ring, including techniques such as agent sandboxing and threat modeling, signal the growing importance of the engineering and oversight needed to manage AI in production. Earlier editions of the Radar traced the growing capabilities of AI in software engineering, and volume 34 examined the friction and risks of using AI at scale. Volume 35 turns to managing those risks as adoption grows.

One indication of this focus is where the Radar makes its strongest recommendations. The 11 practices and technologies placed in its Adopt ring, including techniques such as agent sandboxing and threat modeling, signal the growing importance of the safeguards needed to run AI reliably in production and underline the need to act now.

In developing this edition, the Technology Radar team observed that as agents are handed larger, longer-running tasks, the hardest engineering challenges increasingly lie beyond interactions with the model itself: securing the environment in which agents operate, deciding what they can access, establishing meaningful human oversight, and controlling the cost of running them.

The release states that not long ago, the question clients were asking was how fast agents could write code. Now, it is whether anyone can trust, review, and afford what they produce. Models will keep improving, but lasting value will go to the organizations that invest in the engineering and oversight around them.

The themes explored in this edition of the Radar include what it takes to build and manage the system around the model. These include harnessing engineering at scale, putting more agents to work without losing human judgment, and the economics of running AI at scale. The release also cautions against defaulting to a frontier model for every task and encourages teams to choose models deliberately based on task requirements, cost, privacy, and operational considerations.

The release states that the environment around an AI agent, including its prompts, tools, memory, guardrails, and feedback loops, is becoming a product in its own right, with owners, versions, and dependencies. As a growing ecosystem of plugins and supporting tools takes shape, organizations must treat the harness as an engineered system, with appropriate evaluation, security, access controls, and ongoing maintenance.

The release also notes that developers are increasingly running multiple agent sessions in parallel and delegating longer-running tasks, including unattended overnight builds. But as agents produce more work, human review and software delivery pipelines can become bottlenecks. The release cautions against "human in the loop theater," where approval becomes a ritual rather than a meaningful check, and encourages teams to assess their review capacity before scaling agent use.

The release states that proprietary frontier models are no longer the obvious default for every task. Open-weight models have narrowed the gap for many workloads and offer additional deployment options, including running within an organization's own infrastructure. The release cautions against defaulting to a frontier model for every task, encouraging teams to choose models deliberately based on task requirements, cost, privacy, and operational considerations.

The release states that the cost of AI extends well beyond model pricing. Routing, batching, storage, networking, and scheduling all influence the economics of running AI workloads in production. Much of the innovation needed to make AI economically sustainable lies deeper in the technology stack, where models, GPUs, storage, and networks are managed as programmable resources through a common platform layer.

Together, these findings point to a broader shift in enterprise AI: realizing its potential requires more than access to capable models. Organizations must build the engineering practices, infrastructure, and human oversight needed to operate agents responsibly at scale.

Thoughtworks, a global technology consultancy that integrates design, engineering, and AI to drive digital innovation, is over 10,000 Thoughtworkers strong across 18 countries. For 30+ years, we've delivered extraordinary impact together with our clients by helping them solve complex business problems with technology as the differentiator.