Explore

Use prompt versioning as the trigger for automated eval CI jobs

Doneyli recommends versioning your system prompts in Langfuse (free of charge) and treating a prompt version change as the trigger for kicking off the fully automated eval: configure an automation that triggers on any prompt changes in the agent config, with the action set to launch a CI job in GitHub. Keeping multiple versions of prompts also lets you track how the prompt evolved over time, and you can promote a previously tested version back into production when regressions are observed in the current version. All of this can be done via the API and CLI instead of clicking through the UI; tagging the promoted version with a production tag kicks off the eval pipeline.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Perform error analysis on traces: check steps, tool calls, and outputs

Doneyli recommends performing error analysis by looking at your traces to understand how the agent is behaving. When reviewing traces, check whether the agent is taking all the steps it needs to take, calling all the tools it needs to call, and giving the output that you want.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Evaluation approaches: LLM-as-a-judge, code-based evals, deterministic checks

Use LLM-as-a-judge, i.e., another LLM checking what the LLM did, as a scalable evaluation approach. Use code-based evals for classification tasks where you can check for an exact match. Use deterministic checks for behaviors like detecting all-caps messages.

endorsed

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Automated scores for compliance, failure modes, and guardrails

Define scores for your use case that are automatically calculated whenever a new interaction or trace is captured by the system. Score whether the information given to the customer is compliant with banking information — e.g., if the customer asks about the weather, that is not compliant. Score for failure modes and for guardrails blocking answers given back to the customer. Use scores to filter traces for specific intents, e.g., filtering for traces where customer intent was investment advice.

endorsed

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Build golden datasets from a maintained spreadsheet or from production traces

Doneyli describes two main ways to create golden datasets: have a team maintain a spreadsheet of things to test on the agent and upload it as a CSV, or add items that are already part of your production traces. He frames the latter as 'dogfooding' the dataset — building it from production traces as long as you've determined they are good examples of desired future outputs.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Evaluate against specific evaluators with minimum score thresholds

Doneyli recommends evaluating agents against specific evaluators such as source recall, language match, correctness, and non-unsolicited upsell, and defining minimum score thresholds for each evaluator: if a score doesn't clear the bar, it's not considered a pass to go into production.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

PII masking before persistence in observability systems

Enable PII identification so any data deemed PII is automatically masked before it gets persisted into Langfuse. Financial institutions do not want to persist credit card information or social security numbers in an observability system; they want that data kept in systems of record they own. You cannot control what your users input into the system, so you must catch sensitive data (like credit card numbers in clear text) when instrumenting and building your evals. Use a masking function that immediately masks any defined PII so it is not persisted in plain text in your system.

endorsed

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Building data sets: inputs, expected outputs, out-of-scope requests, stable benchmarks

Data set items should contain the input plus the expected output, which can be a sample reply, expected behavior, or keywords to be covered. Include out-of-scope requests in your data set that you want the agent to correctly defer, such as refusing to give investment advice or answer weather questions. Use data sets as a stable benchmark, either for offline iteration or as part of your CI/CD.

endorsed

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Curate multiple datasets per use case and technical requirement

You typically end up with different datasets to measure specific things or specific areas of quality in your agent fleet. Maintain a red-teaming dataset that looks for prompt injection attacks and any attack that would allow someone to circumvent guardrails and obtain information they shouldn't get. Maintain a dataset specific to a use case, e.g., knowing which cases have been disputed and whether the dispute has not been solved. Curate datasets that are specific not only to the use case but also to the technical requirements you have for that agent.

endorsed

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Datasets as test cases: evaluation criteria anchored on expected outputs

Build evaluation criteria anchored on datasets: a set of questions and expected outputs that allow you to measure the quality of outputs of your LLM or AI solution. Think of datasets as your test cases if you come from a software development background. Do not ship software into production without passing it through QA — merging straight into main or prod is not a best practice. Datasets give you the evidence that you passed the agent through a series of steps or requirements that measure all the things you want to capture. The expected output in a dataset does not necessarily need to be the answer the agent will give; you can measure the intent of the agent, what the agent is supposed to do, or behavior.

endorsed

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Ask SMEs for raw commentary, then use an LLM to extract failure-mode categories

Doneyli recommends not asking the SME to invent categories — instead ask for unfiltered raw commentary on the traces. Then use an LLM to find all failure modes or categories that are patterns across the expert's comments, because LLMs are great at pattern matching and extracting information from text. This keeps it simple for the domain expert and uses LLMs for what they are great at.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Blend LLM-as-a-judge and deterministic evaluators for a balanced mix

Doneyli recommends measuring agent outputs across both LLM-as-a-judge scores/evaluators and deterministic evaluators, and building a library of evaluators so incoming traces automatically get their scores calculated — or alternatively having scores calculated directly from the client application in real time and surfaced back to the system. The more balanced the mix, depending on the categories you identify, the better: if you rely 100% on LLM-as-a-judge, there's a non-zero chance some outputs will be inaccurate because that's the nature of LLMs. Deterministic checks always produce exactly the same output for the same inputs — a code evaluator gives the same output 100% of the time for the same input, not 95% or 80% — enabling a blended assessment of agent quality.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Regulated industries: validated ground truth and human-in-the-loop

In regulated industries, make sure you have a ground truth data set that is validated by your customers. Run human-in-the-loop systems when shipping agents in regulated industries. Run a fully human-in-the-loop system first, then over time slice the data into parts you are confident to automate versus parts where you always want a human in the loop.

anecdotal

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Attack acceleration first and use more steps in the first 5-10 meters

The host recommends attacking acceleration first when learning to sprint, since most sports rarely express max speed (only wingers covering 40-50m breaks benefit). Taking huge long strides in the first 5-10 meters prevents generating enough horizontal force; athletes need more steps (higher cadence) in the initial phase instead.

Host's endorsed heuristic with a stated exception for wingers

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

Start returning runners with baseline plyometrics and short-distance running

For someone who hasn't run in years, begin with very basic baseline plyometrics and keep all running to very short distances.

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

Three acceleration fixes: stay forward, higher cadence, stay on train tracks

Fix 1: stay forward enough for longer — standing up immediately after initiating a sprint means you can't get enough horizontal force into the ground. Fix 2: take more steps in the initial 5-10 meters instead of huge semi-lunge strides, even though long strides feel like pushing hard. Fix 3: keep on the train track — avoid lateral sway and driving out to the side each step, especially for bigger athletes.

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

Feedback policy for training cells: synchronous/asynchronous stimulation scaled by critic-predicted surprise

If the cells do something bad in the game, give them asynchronous stimulation; if they do something good, give them synchronous stimulation. Use the critic in the silicon decoder to predict surprise — figure out how far the action is from the silicon critic and how surprising it is — then scale feedback based on that surprise. If the critic expected an action to be good and it turned out really bad, the cells receive negative feedback. Scaling feedback by surprise modulates it in a more equal way so the system doesn't keep receiving negative feedback and can build connections over time. Positive and negative feedback are defined in relation to the game — if it does something bad in the game, asynchronous feedback is sent to reinforce not doing that action.

Method, not a verified protocol.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Strength-to-bodyweight benchmarks: red flags and green lights

For a contact field sport athlete, a back squat below 1.5 times bodyweight is a major red flag indicating non-functional strength — concerted effort should be put in right now, even if playing the season or pushing other goals is somewhat negative long-term. For a front row forward in rugby, not being near a double bodyweight back squat is a much bigger issue than for a winger. Once an athlete gets beyond double bodyweight back squat for many field sports, they can put the brakes on strength work a little bit. Going from a 200 kg back squat to 220 kg as a 100 kg player is beneficial, but probably not as beneficial as going from 180 to 200 was. The speaker uses three brackets when evaluating athletes on day one — really unhappy, pretty happy, very happy — based on strength-to-bodyweightrat

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

Olympic lifts: practice regularly, prioritize technique over load

Practice power cleans (or hang power snatches, block snatches) regularly and make progress — that alone gets most of the benefit without maximal loads. The same power clean weight performed with sharp technique (contract-relax sequencing, catch at parallel) is far more beneficial for sport than the same weight with poor technique; ripping the bar off the floor and yanking it to the shoulders turns the lift into a poor upper/lower body concentric exercise with minimal sport transfer. Push top-end weight in the off-season when it makes sense — going from 0.8x to 1.1x bodyweight power clean year-over-year is fantastic progress — but don't chase 170-190kg cleans unless you are a thrower or American football player.

Host's endorsed position; some numeric figures flagged as possible numeric errors

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

Get strength, body composition, and conditioning first; Olympic lift ratios are secondary

Get the strength numbers, the required body composition, and conditioning dialed in first — then Olympic lift ratios become secondary. As a composite ceiling, an athlete with a 2.5x bodyweight back squat, a 3-minute 100-cal air bike, and a 1.5x bodyweight power clean is the best F1 car, though most athletes cannot reach this. In American football, key metrics include clean, back squat, single-leg squat, bench, vertical, flying 10, accelerating 10, 40, and shuttle — every one plays a big role. For a 90kg athlete back squatting 135kg (1.5x bodyweight), a single-leg squat target is roughly 60-80kg (about half the back squat load).

Host's framework; several numeric figures flagged as possible numeric errors, and the speaker's own math was uncertain

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength