Thoughtworks, a Chicago-based technology consultancy, is shifting its guidance from what artificial intelligence (AI) agents can do to how companies keep them secure, supervised and affordable as the tools take on larger, longer-running tasks.
In volume 35 of its Technology Radar, published on Wednesday, the consultancy said the hardest engineering challenges increasingly lie beyond interactions with the model itself. They include securing the environments agents run in, deciding what they can access, keeping human oversight meaningful and controlling running costs.
“Not long ago, the question we heard from clients was how fast agents could write code. Now it’s whether anyone can trust, review and afford what they produce,” said Rachel Laycock, chief technology officer at Thoughtworks. “Models will keep improving, but lasting value will go to the organizations that invest in the engineering and oversight around them.”
The consultancy, which integrates design, engineering and AI, publishes the Radar twice a year based on its work with clients.
The latest edition places 11 practices and technologies in its Adopt ring, the Radar’s strongest recommendation. They include agent sandboxing and threat modeling, which the report said signal the growing importance of safeguards for running AI reliably in production.
Systems around agents
The report said the “harness” around an AI agent, including its prompts, tools, memory and guardrails, is becoming a product in its own right, with owners, versions and dependencies. As plugins multiply, it said companies must treat it as an engineered system, with evaluation, security, access controls and maintenance.
Developers increasingly run several agent sessions in parallel and hand off longer tasks, including unattended overnight builds.
But as agents produce more work, human review and software delivery pipelines can become bottlenecks. The Radar warned against “human in the loop theater,” where approval becomes a ritual rather than a meaningful check, and urged teams to assess their review capacity before scaling up agent use.
It also said proprietary frontier models are no longer the obvious default, as open-weight models have narrowed the gap for many workloads and can run in-house. It urged teams to weigh cost and privacy when choosing models.
The cost of running AI also extends well beyond model pricing.
Routing, batching, storage, networking and scheduling all shape the economics of production workloads. The report said much of the innovation needed to make AI economically sustainable lies deeper in the stack, where models, graphics processing units (GPUs), storage and networks are managed as programmable resources through a common platform layer.
Thoughtworks concluded that realizing the potential of enterprise AI will take more than access to capable models. Companies must build the engineering practices, infrastructure and human oversight needed to operate agents responsibly at scale, it said.