IronBee builds a browser agent that skips the LLM and checks the order itself

The demo is small and specific. According to IronBee's own blog post, Express signed in, added a product, entered an address and a card, and waited until the order page read COMPLETED: nine actions, 6.7 seconds.

The decision engine behind it is called Jev, a model from TypeSafe that answers typed multiple-choice questions. IronBee says there is no LLM in the loop. Every step is one call to Jev, which returns the option it picked, a probability for each option and a confidence score. No text is generated and no screenshots are read. A decision takes about 300 ms from Turkey, with most of that the trip to the server. The cost figure is what the company is selling. Input runs at $0.042 per million tokens and output is free. IronBee says that brought a full replayed checkout, judging and review included, to $0.00054. The answer can only be a key from the list the harness sent, so model output never becomes a selector, a coordinate or a script. A run can only click what was already on the screen.

That is a narrower bet than the screenshot-driven agents shipping elsewhere. It is worth watching whether the approach survives pages that do not offer a clean numbered list of controls.

The failure that sold the design

Speed is easy to film. The more useful run in the post is a checkout that went wrong. The page displayed "Order placed successfully!", and the order API returned 200, but the payload inside that response said FAILED. IronBee's agent flagged the run as failed in 7.57 seconds across 8 actions and 9 decisions. It pointed at the API response and then at the backend log line: notification-service, Order #131 Could Not Be Processed, Reason: Insufficient inventory.

The page said "Order placed successfully!". The order API answered 200, but the order inside it said FAILED.

Nobody wrote an assertion for any of that, the post says. The judging step is itself a Jev question: did the goal hold, and what evidence supports it. All the numbers in the write-up come from IronBee's own runs between 21 September and 1 October 2026. The first version ran on 24 September. It signed in and added a product to the cart 3 times out of 3 in 2.85 to 3.05 seconds.

That caution matters, because most of the browser-agent material published this week is not a neutral benchmark. A 400-line reference implementation from xhluca shows the opposite architecture. It gives Gemini 3.8 Flash screenshots and browser controls in 394 lines of Python, including a recorded run that books a workshop and verifies the confirmation. Both approaches claim to check the outcome. Only one of them pays per screenshot.

Verification is the theme of the week

The same argument shows up in coding tools. Dotpals, an open-source desktop companion for Claude Code, Codex and Cursor, reads test results from command output rather than the agent's word. It sends the agent back when tests fail, and it also catches a test edited just to make it pass. Ian Maurer's write-up cites the OpenClaw team deleting around 400,000 lines of their own tests with little change in coverage. It argues for outside-in tests that state user behaviour instead of a growing pile of unit tests. Threadnote ran a small paired study on five public repositories. It reports 65.62% fewer lifecycle tokens per verified completion and 46.14% less time with a stored handoff, passing verification on 5 of 5 tasks against 4 of 5 for files alone. The company is explicit that this is its own single-model study on build 5.1.0-beta.2, not an independent evaluation.

Its line on the problem is blunt: "Done" is a sentence an agent can generate, and it is not a test result.

Durability is getting attention too. Pi Durable, an experimental framework from Arendelle released alongside Pi 1.0, checkpoints conversations so a new process can resume them. It only replays an interrupted tool call if the tool declares replay safe, because re-sending a payment twice is not the same as re-running a search. Ostoyae, a graph runner for coding agents, runs each job in its own git worktree. It only counts work once a check command passes on a fresh checkout.

Why enterprises stall

The gap between demos and deployments is measured. An MIT Technology Review Insights survey of 300 data and technology executives found that on average only about 34% of organisations' agentic AI projects reach production. The survey was produced by its custom content arm rather than its editorial staff. Data fragmentation was the most cited obstacle to expanding agents' access to knowledge, at 55%.

On the research side, a paper revised on 1 October introduces StructMemEval, a benchmark for whether agents organise long-term memory rather than just recall facts. Its authors report that retrieval-augmented LLMs struggle with tasks like transaction ledgers and to-do lists. They add that modern models do not always recognise the required structure unless prompted.

None of this makes a 6.7-second checkout the hard part. IronBee's own post spends most of its length on things that broke. Its central claim is narrow: pick from a list, check the pick, and judge the run against evidence instead of the page's success banner. That is a smaller promise than an agent that reads the web, and a more testable one.