Delineate deterministic steps from agent steps in the run chain
In the recorded run chain, clearly delineate deterministic steps (code-based fetch or match) from agent steps (extraction, conclusion).
42,607 results · Most relevant
In the recorded run chain, clearly delineate deterministic steps (code-based fetch or match) from agent steps (extraction, conclusion).
Rerun individual recorded steps rather than the whole run to isolate exactly where a result diverged between points in time.
Use Skill Doctor, a skill you can run on your single-player setup that crawls through all of your agent conversations in Claude, Codex, Pyre, whatever else, and does a one-time grading. If you haven't done this grading approach before, run it across your setup and see what happens — it's a first go to see what you can optimize right now and get some low-hanging fruit. It grades in batches for a little bit of token efficiency and produces a skill report card showing how the agent performed across all defined rubrics. While grading should probably be continuous rather than whenever someone remembers, a one-shot run is a nice starting point.
Consequential creator advice requires independent verification before Today can promote it.
When you're using scoring, have an actual dashboard that explains how all of the conversations went. Track over time how the efficiency score has been fluctuating as agents constantly score setups. Attribute changes you made in your agent setup to inflections in the score to get proof points that the change was real — for example, when you merge a skill the self-improvement agent suggested, mark the inflection point on the score dashboard to see that efficiency went up by a certain percentage.
hedged
Automatically trigger an agent to review a conversation across your grading criteria when a PR is merged, and grade conversations as soon as they happen. You can also set up a teacher agent to go through the rubric and decide whether the conversation gets good marks on efficiency or is below the bar.
Give agents upfront instruction on how to test changes so they don't waste time exploring the test environment, and instructions on finding a port number first and then going for it, to save efficiency in tool calls.
Paddy Pimblett says Cage Warriors is the best show in Europe by a mile, and that if you want to get into the UFC and you're in Europe, you need to have Cage Warriors. He encourages fight fans to get UFC Fight Pass to watch who's coming up, saying the level of Cage Warriors' top guys is UFC level.
Pimblett recommends hiring a nutritionist and S&C coach to manage weight cuts and food (Paul Reed weighs out his food, tracks calories and sodium), doing plyometrics as part of training, taking creatine, lifting weights twice a week to build muscle mass, and getting a DEXA scan to track body composition changes — he says without Paul Reed they'd still be in the dark ages on making weight.
hedged
Philipp Schmid recommends that with managed agents, you can bootstrap the agent's computer environment by inlining skills and AGENTS.md files directly, providing a GitHub repository (private or public) with your skills, or connecting a GCS bucket.
Philipp Schmid recommends using Workspace's MCP server for Gmail and spreadsheets to integrate those tools with an agent, and notes that calling your own managed agent uses the Interactions API create call with the agent ID instead of a model ID. He also notes that if you don't have an MCP server, writing a local function is also possible.
Philipp Schmid recommends keeping system instructions very concise and short, because system instructions are included in every request and long ones make all requests much more expensive. He also recommends using skills with very short descriptions that give the agent guidance on when it can use that skill, and not creating super long, dense skills, because that's also a high cost — context is very valuable, so decouple and separate content into skills/slices with dedicated references and scripts for when the agent needs more context.
Philipp Schmid recommends having the agent create a function call and sending back the function result yourself when you want certain executions to run outside the agent sandbox.
Philipp Schmid recommends setting up a cron trigger to schedule the agent to run a certain input in the same environment on a recurring schedule, for example for due diligence or auditing, and also scheduling working hours specifying when agents can come in or have to work.
hedged
Sid Pednekar advises that depending on adoption stage, organizations can deploy different types of team agents: expert agents acting as subject matter experts, workflow agents operating across streams such as Slackbots that create, open, and close tickets, and chief-of-staff agents that keep tabs on the state of the organization and what needs to be done.
After assembling the right people and knowledge, Dr. Pedneker recommends asking how to make that knowledge reusable and accessible. Depending on their adoption stage, organizations can deploy different types of team agents: expert agents acting as subject matter experts, workflow agents that work across streams (such as Slackbots that create, open, and close tickets), and chief-of-staff agents that keep a tab on the state of the organization and what needs to be done.
Dr. Pedneker recommends getting the right people in the room by breaking down organizational silos — the practice leaders lean on. Organizations can create forward-deployed engineers, forward-deployed product managers, and the inverse — forward-deployed portfolio managers.
Chuck recommends that instead of opening a book or a how-to manual — which he says would take 20 years to accomplish what you could in one year — you sit down with an AI agent and say: this is what I'm thinking, tell me if this would work, and tell me what resources I need to find to make this happen. He adds that you'd be amazed at the answers you get.
Chuck advises asking your AI agent: what do I need to do to grow this business? He pairs this with the observation that none of us have all the ideas or know how to do it.
Chuck wants every person on his team in Lattato every day. Staff have to log into the hub, where company messages written by Latido are posted, and that is where they get their tasks.
Chuck advises allowing the AI agent to take all your data and give you a blueprint on how to accomplish your goals; you can then decide how you want to accomplish the blueprint. After getting it, you can just go and tell the agent to do it.
Choose sources → to get a library of your own and a daily brief.