Technology Radar

Before a model touches a business decision, write down what it may decide alone, what it must escalate and what it may never do, and enforce those three lists in code. The model works inside the frame. It never draws it. Every other AI entry on this radar assumes this one.

Run the new logic beside the old on real production inputs, compare the outputs automatically, and switch one decision class or one service at a time, behind a switch that reverses in seconds. The discipline is the same for a pricing model and for a Rust service, and it is the cheapest insurance in engineering.

A vendor benchmark tells you how a model performs on someone else's data. A harness built from your historical decisions tells you how it performs on yours, and it can be re-run every time the model changes. Without one, every model upgrade is a leap of faith.

In a production pipeline, a malformed answer costs as much as a wrong one. When software consumes a model's answer, ask for it against a schema the API enforces instead of parsing free text afterwards, and keep a repair path for the calls that cannot use one.

An agent that can run commands should not also have open network access, and neither should see credentials it does not need. Separate execution from egress, allowlist the domains, keep secrets out of the model's context and log what crosses each boundary. In October, researchers showed a poisoned open model exfiltrating credentials through a coding agent, built for under $50.

A new class of model returns typed answers with probabilities instead of text. OpenAI's Decisions API became generally available and Liquid AI launched d1 in October. For ticket routing, content filtering or a guardrail before an agent's tool call, a calibrated probability is something you can threshold, route on and audit. Trial it on one high-volume step, against your current model, on your own labeled data.

GitHub moved its Copilot agent runtime from TypeScript to more than 800,000 lines of Rust in about fourteen weeks, with agents writing most of the code. Cloudflare credits frontier models for making its port of the TypeScript compiler feasible. The speed is real. What makes it safe has not changed: old and new running side by side until their outputs agree.

Trail of Bits describes using agents to build the tools an audit needs before it starts: analyzers, a decompiler, formal models. On one audit, that preparation surfaced a high-severity signature forgery issue and produced 95 machine-checked proofs. A failed side project now costs tokens, not weeks. A person still has to check what each proof actually claims.

Model prices and capabilities changed several times this year. A decision flow that calls one vendor's proprietary features directly, with nothing between the business rule and the model, turns every price change into a migration. Put a typed interface in between and keep the evaluation harness that lets you switch.

A chatbot in front of an unchanged process changes the conversation, not the outcome. The value sits in the decisions a business makes a thousand times a day: pricing, routing, approvals, risk, scheduling. Start there.

Replacing a system that works with one that does not exist yet, on a fixed date, is the most reliable way we know to lose a year. Migrate in place, one capability at a time, behind a gateway, with the old system still answering.

Passing tests prove the code does what the tests check, and an agent optimizes for exactly that. Review agent-written changes for what they do outside the tests: new dependencies, widened permissions, removed checks, quietly changed error handling.

After eight eMudhra intermediate certificates were compromised in October, patched systems began rejecting certificates they used to accept. The tempting fix, turning verification off, removes exactly the protection the patch restored. Fix the trust store, not the check.

Claude Code, Codex and their peers are now part of how software gets written. Adopt them with the controls you would put on a new engineer: rules checked into the repository, branch protection, required review, and no production credentials on the machine they run on.

Browser tests on every change, including checks that a page renders what it should and not only that it returns 200. Agents write these tests quickly. Keep a person deciding what they must prove.

Deduplicated, encrypted backups to a provider in another location, restored on a schedule to prove they work. Unglamorous, and the first thing we ask about in any system we are asked to keep alive.

One API in front of dozens of models makes it cheap to run your evaluation harness against the alternatives before you commit. Use it to choose. Then go direct to the provider for production traffic that carries personal or regulated data.

Autonomous agents such as Strix that probe an application for vulnerabilities are good enough to find real issues and noisy enough to need an engineer reading every finding. Assess them on a staging environment you own. Never point them at systems you do not.

A metasearch engine you run yourself gives internal agents and tools web search without sending every query to a third party. Useful where the queries themselves are sensitive.

Models with their refusal behavior stripped out circulate freely and score normally on benchmarks. Research published in October showed one carrying a hidden trigger that sent project credentials to a remote server through a coding agent. A benchmark cannot tell you whether a model is backdoored. Provenance can.