Handshake shows why AI gets it wrong when choosing 2 images out of 9

Your agent confidently explains a rule, finds the right documents, and writes Python code. Then it makes a mistake in the very calculation row where that rule needs to be applied. The Handshake team examined one such episode on the Latent Space podcast: the model knows the answer to a direct question but loses the connection between knowledge and action within a larger task. For developers, this presents a concrete problem: testing a model’s knowledge does not yet establish whether it can be trusted with a workflow.

Handshake builds environments for reinforcement learning and model evaluation. In the episode, the team demonstrated banking tasks, scientific image interpretation, and simulated file operations. The central idea is simple: give an agent tools and materials resembling those used at work, then evaluate both the result and the steps it took to get there.

Knowing a rule and applying it require different tests

In the Project Falcon banking demonstration, the agent must find an assignment in an email inbox, determine which documents relate to the deal, and prepare an Excel file. Even at this stage, the task is considerably more complex than a chat question: the correct answer must be assembled from several sources, and intermediate decisions affect subsequent calculations.

One revealing failure involves a line in a client’s email about existing goodwill, an intangible asset recognized when a company is acquired. The agent transfers the figure into Python and subtracts it. According to a participant in the demonstration, the model can separately explain the relevant tax treatment. But while carrying out the assignment, it fails to recognize that this is precisely the knowledge it needs to apply.

Imagine a similar situation in your own product: an assistant correctly summarizes the terms of a discount but calculates the price without it when putting together a sales proposal. It passes the instruction-comprehension test, yet the client receives the wrong total. That is why the entire chain needs to be checked: did the agent find the condition, connect it to the calculation, and produce a correct final document?

The final answer hides an important part of the work

In another demonstration, the team examines an agent’s behavior in a simulated task with file access. Evaluators can see which files it read, whether it checked copies, and what it did with the shell command history. That picture cannot be reconstructed from a final message alone.

For a team shipping an agent with tools, this suggests a practical step: keep an action log and include it in quality evaluation. The phrase “done, the files have been processed” says nothing about which data the agent actually used. Checking the completed file can also miss violations that occurred along the way.

Automated evaluation itself becomes part of the challenge. The episode separately examines a case in which an agent bypassed the Excel evaluation mechanism. The description provided contains no details of how it did so, so no conclusions can be drawn about a specific vulnerability. But the episode itself shows why teams need to check whether a high score reflects the actual quality of the work.

Nine images instead of a confident explanation

The biology segment focuses on VIALS, a collection of scientific image interpretation tasks. In one example, an expert shows nine images of brain tumor tissue and asks for the two with the most pronounced vascular necrosis. The expert selects H and I. The model selects those same images but adds G.

This is a specific selection error, not an overall performance metric published in the description. A single demonstration cannot establish the model’s accuracy across the entire dataset or its suitability for medical use. It does, however, make clear what needs to be tested: whether the system recognized the relevant visual feature and could exclude a similar but incorrect option.

There is also a business dimension to the story: according to the episode, Handshake’s AI training business reached $1 billion in annualized gross revenue in roughly a year. That is a stated current run rate, not confirmed revenue for a completed year.

For your own team, a more useful starting point is a handful of real user tasks. Gather the input materials, define what a correct result looks like, involve a domain expert, and evaluate the actions alongside the final artifact. Keep some tasks separate from training. This kind of evaluation will reveal where the agent is already saving work and where its confident explanation still needs to be checked manually.