Thoughtworks published Technology Radar Volume 35 on November 4, and the headline isn’t a model release or a new framework — it’s a mandate. Agent sandboxing has moved from Trial to Adopt. Threat modeling follows it. And Thoughtworks gave a name to the failure mode quietly spreading through engineering teams everywhere: human-in-the-loop theater. With 117 blips and 87 new entries, Vol. 35 is less about what’s emerging and more about what you should already be doing.
Sandboxing Is the New Baseline
Until now, Thoughtworks recommended sandboxed execution specifically for coding agents. Volume 35 drops that qualifier. Every agent — coding or otherwise — should run in an isolated environment with restricted file system access, controlled network connectivity, and bounded resource usage. That’s the Adopt call, which in Radar terms means “do this in production.”
The practical tool for this is mise, which also moves into the Adopt ring this volume. It consolidates tool versioning, environment variable management, and task execution into one binary — and critically, it works identically for human developers and CI/CD pipelines, so the environment an agent runs in matches what your team actually uses.
# Install mise
curl https://mise.run | sh
# Pin versions — identical across human sessions and agent runs
mise use node@22 python@3.12 --globalThe rationale isn’t just security hygiene. Sandboxed agents are easier to debug, easier to audit, and when something goes wrong — and it will — their blast radius is contained. This should have been standard practice a year ago. Vol. 35 is saying you don’t have a good excuse anymore.
Name the Theater, Stop the Theater
The sharpest insight in Vol. 35 is a phrase: human-in-the-loop theater. It describes what happens when developer teams run multiple parallel agent sessions, the volume of changes outpaces their ability to review meaningfully, and approvals become rubber-stamps. Oversight exists in name only.
Thoughtworks doesn’t pretend this is avoidable as agent usage scales — they offer a structural fix instead. Risk-tiered human oversight (in the Trial ring) aligns review depth with what a change could actually break. High-impact, high-risk changes get real human judgment; low-risk, well-tested changes get automated gates. The goal is to reserve human attention for decisions where it creates value, rather than spreading it thin across everything.
If your team is approving agent-generated pull requests at a rate that makes real review impossible, you’re already doing theater. The fix isn’t to slow down agents — it’s to design a system where your approval still means something.
The Harness Is the Product Now
Thoughtworks formalized something the industry had been circling around: the concept of harness engineering. The formula is “Agent = Model + Harness” — the harness being everything that constructs the environment around the model: memory, context, feedback loops, guardrails, tools, and the quality gates that determine when output is acceptable.
Developers who think the job is selecting the right model and writing good prompts are thinking about half the problem. The harness is what makes an agent reliable in production, and Vol. 35 reflects a maturing ecosystem of tools for building them. DeepSeek Harness (DSH) gets a nod here — an open-source MIT-licensed runtime where every model interaction is captured in an append-only session log. Runs can be inspected, forked, and replayed. That’s the kind of observability a production harness needs.
Cost Is an Infrastructure Problem
The fourth theme in Vol. 35 is the least comfortable one: running AI at scale costs more than model pricing. Token economics, Thoughtworks argues, ultimately comes down to infrastructure economics — GPUs, inference engines, storage, and networks managed as programmable resources.
Two Assess ring items point at this directly. Context profiling tracks whether your orchestrator’s context window is growing with every edit — a signal that your agents are getting more expensive per run without you noticing. Self-improving harnesses take the next step, using agents to analyze their own execution traces and propose optimizations to the system around them.
Neither is Adopt-ready yet, but the trajectory is clear: cost optimization is moving from a cloud billing concern to an engineering concern.
What to Do This Week
The Adopt ring from Vol. 35 gives you a concrete starting point:
- Sandbox every agent — not just coding agents; use mise for reproducible isolation
- Add mutation testing — especially if agents are writing your tests; verify those tests actually catch bugs
- Run threat modeling on your agent’s access scope — what can it read, write, call?
- Audit your approval process — if you can’t describe what you’re checking, it’s theater
- Try OpenCode as a harness reference — Thoughtworks keeps it in Adopt as the open-source benchmark
Volume 35’s through-line is straightforward: agents are production systems now, and they need to be treated like it. Sandboxing, oversight, cost visibility, and harness engineering aren’t advanced topics — they’re the basics that the industry has been deferring. The Radar says the deferral is over.
The full Thoughtworks Technology Radar Vol. 35 is available online with all 117 blips across the four quadrants.
