Explore

Agent reflection internals: clustering, LLM diagnosis, merging, ranking, scorecard

After getting the judge score for each dimension, the next step is to run the agent reflection process to find the root cause of a low score. Agent reflection internally has four mini-steps. First is clustering: group failing rows (users and traces) by the failing dimension — pure clustering without any LLM calls. Second is the LLM judge (L2 judge): within each cluster, run the LLM to diagnose each row, assigning a root cause plus a tag. Third is merging issues: because the LLM root causes are free-form tags, similar tags that describe the same thing need to be deduplicated/merged to produce canonical root cause labels — for example, one row's root cause says the member's profile data is missing while another says the profile data doesn't show the full picture of the user's browsing; both

Method, not a verified protocol.

AI Engineer · An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn

Set the preacher curl bench angle high enough and press the triceps into the pad

The host discourages setting the preacher curl bench angle too low, because it creates a tendency to sit down, and discourages failing to press the triceps fully against the pad, which leaves the biceps vulnerable to tendon tears and shoulder stress in the lengthened position.

ATHLEAN-X™ · The ONLY 2 DB Biceps Exercises You Need (NO, SERIOUSLY!)

Engage lats by pushing back into the preacher pad

The host recommends pushing back into the preacher curl pad so hard you could almost pull the bench toward yourself, engaging the lats to support the shoulder girdle and protect against a torn bicep tendon and shoulder stress; leaning into the bench over the top keeps it from moving.

ATHLEAN-X™ · The ONLY 2 DB Biceps Exercises You Need (NO, SERIOUSLY!)

Avoid a 45-degree bench angle for the incline drag curl

The host discourages reclining the bench to 45 degrees for the incline drag curl, because it puts the arms too far behind the body and places unnecessary stress on the bicep tendon and anterior shoulder structures; 60 degrees is preferred.

ATHLEAN-X™ · The ONLY 2 DB Biceps Exercises You Need (NO, SERIOUSLY!)

Approach from a good distance with good delivery, tonality, and speed

The creator advises approaching a woman from a good distance so she feels safe and comfortable, using good delivery, tonality, and speed; he claims most women will then be very appreciative and grateful for the compliment, even if they have a boyfriend.

Coach Kyle · STOP Making This Mistake When She Says "I Have a Boyfriend"

If she says she has a boyfriend, smile, say 'no worries, take it as a compliment,' and walk away

The creator advises that when a woman says she has a boyfriend, the man should smile, tell her 'no worries, take it as a compliment,' and then walk away gracefully.

Coach Kyle · STOP Making This Mistake When She Says "I Have a Boyfriend"

Pull future capabilities into today's product with engineering, then delete the scaffolding

Roman recommends seeing where the world is going three to six months out and pulling that capability into today's product, even if it requires a lot of engineering or changing the product so users can interact with the new capability. Then, once models are good enough natively, delete the scaffolding — and repeat the cycle, which he credits with building user trust.

hedged

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Delete the product: remove scaffolding and product overhang as models get smarter

Roman recommends 'deleting the product' — removing scaffolding and product overhang as models get smarter — and being comfortable making hard decisions that might upset a small set of users in order to keep the product simple and powerful.

hedged

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Keep the editable policy surface small so the auto-research layer cannot rewrite the whole codebase

Irina recommends keeping the editable policy surface really small: the auto-research layer should not have permission to randomly rewrite the whole code base. This bounds the blast radius of automated changes.

AI Engineer · Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati

Preserve information provenance: first-hand, second-hand, verified, or uncertain

Erina Karati recommends that agents sometimes need to know where information came from — preserving whether it was first-hand, second-hand, verified, or uncertain. This addresses the rumor-hardening failure mode directly.

AI Engineer · Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati

Give agents a goal.md plus agents.md rules file and a scratch pad for notes

Elie Bakouch recommends using a goal.md file plus an agents.md defining the rules so agents know the objective and constraints, letting them propose ideas and submit jobs, and giving them a scratch pad as active memory where they can write notes and decisions. He notes the harness could have been replaced with a simple /goal command since goal.md plus rules is sufficient scaffolding.

AI Engineer · We Let Claude Code and Codex Race Human Researchers — Elie Bakouch, Prime Intellect

Define your accessibility requirements before having your agent write code

Jemima Abu recommends defining your accessibility requirements before having your agent write code, since every organization has different requirements depending on their field, and knowing what the accessibility standards are for your app and keeping that in mind when prompting your agent.

AI Engineer · Teach Your AI Agent to Build Products Everyone Can Use - Jemima Abu, CAIS

Design specialist tools with complexity intact

The narrator recommends designing tools for specialists and domain experts with their complexity intact rather than simplifying for non-experts, and testing changes with the actual specialist users who will operate the tool.

NNgroup · Complex vs. Complicated

Verify identity before account access, and confirm critical dates and financial amounts

Jesse Hall advises putting explicit checks in place: verify the caller's identity before an account is accessed, and confirm critical dates and anything financial (money) — including the full amount with cents — before charging, since failing these leads to wrong payments or reservations.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Keep the agent listening while slow tools run, and let it multitask

Jesse Hall advises that when a tool runs slowly mid-call, the agent should keep listening so the conversation continues instead of going silent — using an async await on the tool call so the conversation and tool stop sharing a thread, handing the mic back to the agent, which can move on and ask another question while the backend tool runs.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Run the complete conversation repeatedly through simulations and turn failures into regression scenarios

Jesse Hall advises running the complete conversation repeatedly through simulations — testing final state, privacy safeguards, unnecessary tool calls, and turn-taking — and when failures are found, converting them into regression scenarios and continuing to test over and over.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Await tool calls asynchronously so the agent keeps talking while the backend works

Await the tool call with async/await so the conversation and the tool stop sharing a thread — this hands the mic back to the agent and lets it keep talking while the backend works. Keep listening while slow tools run, and keep the conversation going. The agent should move to the next piece of information immediately rather than waiting for the backend to confirm each save.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Score agents on the steps they took, not just task completion

Jesse Hall advises that scoring an agent only on task completion lets agents game the metric, so verify the intermediate steps: treat taking 10 steps where 2 would do, or hallucinating a function call, as failures in scoring.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Let the thumb tilt up slightly to create wrist gapping and better alignment

Jeff Cavaliere recommends getting a little gapping in the wrist by letting the thumb tilt up just a bit when pressing. With the thumb tilted up, the carpal bones sit down on the base of support through the radius and ulna, better mechanically aligned because there is space and no compression. He frames this as the best wrist position for dumbbell pressing, and suggests a setup cue: say to yourself to let the thumb tilt up just a little bit to create gapping.

Consequential creator advice requires independent verification before Today can promote it.

ATHLEAN-X™ · Please STOP Gripping Your Dumbbells Like This!

Jump found pull because financial advisors must take notes (or risk jail) and generic AI note takers weren't SEC-compliant

Financial advisors face an unavoidable, compliance-driven project (meeting notes into CRMs), dislike doing it, and generic AI tools were unusable due to unclear SEC compliance — so when Jump showed up, people ripped the product out of their hands.

Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)