Explore

Use a separate supervising agent with separation of concerns: it decides what the worker agent does but does no work itself

The creator recommends that there should be a separate supervising agent that decides what the worker agent gets to do, with separation of concerns: the supervisor can't do any of the work but decides what the worker does. Dick similarly recommends having a supervising agent that runs your policy and is independent of the agent doing the work. The creator gives examples: a request to read someone else's email gets caught in deterministic policies right away — it's not vague, and good controls exist for it. But telling an agent to forward the CEO's email to the outside world is vague, and ideally a supervising agent would catch it and say that's a bad idea and no, you can't do that.

AI Engineer · An AI Agent Got Its User Banned From Resy — Keycard, JPMorgan Chase & AAuth

Keep the auth log outside the agent's hands and brains; maintain a record of who told the agent to do what, for what purpose

Ian recommends that the auth log has to be outside the primary hands and brains of the agent because you can't trust the agent itself. He also recommends maintaining a record of who told the agent to do what thing, for what purpose, and whether that is still true — independent of enforcement points or vulnerability patching — and having the ability to answer the liability question of who did it, which is where the tamper-resistant auth log comes from.

AI Engineer · An AI Agent Got Its User Banned From Resy — Keycard, JPMorgan Chase & AAuth

Sandbox the agent's brain with tools inside; egress proxy as the authorization boundary

Put the agent's brain (the model/LLM) in a sandbox with the tools inside so it can't get API access directly, and place an egress proxy outside the sandbox as the authorization boundary — it does translation between systems, hosts OAuth, and applies authorization at the boundary.

AI Engineer · An AI Agent Got Its User Banned From Resy — Keycard, JPMorgan Chase & AAuth

Switching models with evals: upgrading to Haiku 5.5

To climb the hill of a capability eval you can work on the prompt again or upgrade the model. You can run the experiment again by passing a new model (Haiku 5.5) and running npm experiment. You can't switch models without evals — evals are what make model switching possible because you're measuring everything. Running the experiment with three repetitions takes 13 minutes. Switching from Haiku 4.5 to 5.5 improved faithfulness more than three times, but it is still terrible — the agent is still making things up all the time. Consistent calls got better, so it is at least making the same call over and over. Upgrading the model also costs four times as much money, and whether that is worth it is a question you need to make for yourself in production.

Method, not a verified protocol.

AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI

Running experiments in Phoenix: datasets, tasks, and evaluators

An experiment in Phoenix is a dataset, a task, and evaluators; the task is a function that runs the agent on each example. Phoenix runs the task on every example, scores every result, and stores everything for comparison. The task hands back the report and research, and each eval gets a small wrapper that picks out the fields it needs (faithfulness gets the research as context). You can run the experiment with npm run experiment, which takes about four minutes, and you can target a specific subset of the agent's behavior or a subtask rather than the whole agent.

Method, not a verified protocol.

AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI

Run the meta-evaluation: judge vs. human labels on the dev split

Run the judge against the humans as an experiment — an experiment in Phoenix takes a dataset, a task to run on every example, and evaluators to score the results. In the meta-evaluation, the task is the judge and the evaluator checks whether the judge's label matches the humans. Only run the meta-evaluation on the dev split, not the test split. Attendees can run it themselves with npm run meta.

Method, not a verified protocol.

AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI

Running evals in CI with vTest and gating builds on acceptance criteria

To create an eval that runs in CI, use vTest, a JavaScript test runner Phoenix can plug straight into. agent.eval.ts looks like an ordinary test file because it is one — a describe block and a test for each query in the regression set (Coca-Cola, Apple versus Microsoft, Rivian, and the fake ZZZQ). Each test runs the agent, logs the report as the output, and runs the same evals written earlier — no new evals are needed for CI. At the bottom of the file you gate the build on acceptance criteria: actionability is a target with the average having to stay 0.8 or better, and you gate on the average score, not a single report, because an LLM judge can be noisy on a single report. You can run npm test like any test suite; the reporter prints a scorecard in about four minutes. In a red build where,

Method, not a verified protocol.

AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI

Mapping eval findings to specific prompt changes

Laurie walks through concrete examples of mapping eval findings to prompt changes: because the judge said 'makes no recommendation', the improved prompt asks for an explicit recommendation; because the judge said 'mentions risks with no evidence', the prompt asks for risks with supporting data; because faithfulness found invented numbers, the prompt says to use only numbers that appear in your research; and because reading traces found the agent trying to save files all the time, the prompt tells it to reply with the report instead. The improved prompts live in prompts.ts, with an improved version of each prompt or the option to write your own.

Method, not a verified protocol.

AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI

Step 7: use trusted evals to improve the agent and prove the fix

Once you have evals you trust, use them to actually improve the agent and prove the fix worked. In Phoenix you can filter evals (e.g., 'not actionable'), select them, and add them to a dataset, then run experiments against only the failing evals instead of the entire set — which saves thousands of tests per run on a decent-sized test set. Every prompt change should map to a finding from the evals, and you should not guess: come up with specific things for the agent to do based on previous eval results.

Method, not a verified protocol.

AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI

Use faithfulness instead of correctness when the judge lacks fresh data

Laurie Voss recommends using faithfulness instead of correctness when the judge lacks fresh data: faithfulness asks whether the report makes sense given the data the agent was given, and can work with exactly the same data the agent was given.

AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI

Run agent-driven error analysis over traces

Laurie Voss walks through the error-analysis workflow: open the coding agent, tell it where Phoenix is, have it read traces, come up with failure categories, count them, and write each category back onto traces as annotations; approve the plugin's tool permissions and let it write notes to disk. His run took about nine minutes. An annotation is a human-readable string attached to a trace.

Method, not a verified protocol.

AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI

Don't write vague criteria — write down exactly what makes a report actionable

Laurie Voss discourages writing 'a good response is helpful and accurate' because a judge doesn't know what helpful and accurate mean in this context, and it can't make a yes or no decision from that. Instead, write down exactly what makes a report actionable: it makes a specific recommendation, it names concrete risks, and it looks forward, not just back. He notes this is where a lot of people don't put in enough effort — the explicit criteria section of the rubric.

AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI

Annotate traces in Phoenix to build human ground truth

To annotate in Phoenix, attendees open up a trace, click the little pencil button, and add annotations. Laurie Voss created an annotation in advance called 'human actionable' that can be set to actionable or not actionable, then quickly clicked through all traces and set it on each one. This is where a domain expert would get involved: throw them at Phoenix with 10, 20, or 50 traces and have them read the outputs and give you a golden dataset. The speaker sat down and labeled all 16 of his reports properly against the same criteria he gave the judge, and those are sitting in the seed project so attendees don't have to do it.

Method, not a verified protocol.

AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI

Build several small evals, not one giant super eval

Laurie Voss advises building several small evals today rather than one giant super eval; cramming every possible failure into one big LLM-as-a-judge eval is unreliable and complicated without solving the problem.

AI Engineer · How to Know If Your AI Agent Is Any Good: Evals With Phoenix & Claude — Laurie Voss, Arize AI

Build a shared domain (plan mode) with the agent first

Francesco recommends first building a shared domain with the agent — calling it plan mode — before anything else, then setting up consistency.

AI Engineer · Stop Chatting With Spreadsheets — Francesco Galletta & Felippe Felisola Caso, Monzo

Design a branching harness where decisions branch out and join back

Design a harness with a branching structure where decisions branch out and join back in, so the model can verify assumptions with you or figure out how things fit into the overall system.

endorsed advice

AI Engineer · Stop Chatting With Spreadsheets — Francesco Galletta & Felippe Felisola Caso, Monzo

Define the outcome and success metric first, then track whether AI supports it

When evaluating AI adoption, first define what outcome you're looking for and how you measure success; e.g. if the goal is less back-and-forth with a BI team, validating a new dashboard, or fewer internal ask tickets because a number looks wrong, track that metric and see if AI can support generating that outcome.

endorsed advice

AI Engineer · Stop Chatting With Spreadsheets — Francesco Galletta & Felippe Felisola Caso, Monzo

Move from skill-based prompts to a separate tool when the use case becomes critical

Move from a skill-based prompt approach to a separate tool when the use case becomes critical enough that the trade-offs of the non-deterministic approach are not acceptable.

endorsed advice

AI Engineer · Stop Chatting With Spreadsheets — Francesco Galletta & Felippe Felisola Caso, Monzo

Never have an athlete do an exercise that causes pain

Grubbs's rule: never have any athlete do any exercise that causes pain; when selecting exercises, ask why you are doing it and whether it is creating pain — if so, you probably shouldn't be doing it.

Garage Strength · How The Best Pro Baseball Players Actually Train w/ MLB Strength Coach Stephanie Grubbs

Produce independently checkable certificates rather than claimed independence

Clark recommends building systems that produce independently checkable certificates so others can verify the evaluation themselves, arguing it would be a real shame if all AI safety were left in the hands of the labs or friends of the labs who merely claim they are independent.

Y Combinator · AI Safety Fact vs. Fiction | YC Paper Club