Explore

Put a ton of time into making docs, Claude MDs, and skill files good

Charlie Holtz recommends putting a ton of time into making docs, Claude MDs, and skill files good, saying he's seen this with all the best builders up close. He offers a mental model: think of the Claude MD or Agents MD like whispering in a new intern's ear every time they start working — you'd put a lot of thought into what you whisper, so do the same for agent context files. The Claude MD or Agents MD is the information loaded into the agent's context every time they start working.

hedged

AI Engineer · Orchestras, Not Factories: How the Fastest Builders Work — Charlie Holtz, Conductor

Use the 'don't beat the market' heuristic to decide when to optimize your workflow

Charlie Holtz recommends treating workflow optimization like the efficient market hypothesis: don't try to beat the market. Unless you have real alpha — information about your users or your code base that the models might not know about — you shouldn't be optimizing your workflow too much. If a workflow approach works for everyone, wait for Anthropic or OpenAI to build it into the default harness. A useful self-check is asking 'why isn't this workflow the default?'

hedged

AI Engineer · Orchestras, Not Factories: How the Fastest Builders Work — Charlie Holtz, Conductor

Give agents opportunities to create more of themselves and to collaborate

Charlie Holtz recommends giving agents opportunities to create more of themselves — for example, giving agents APIs to spawn themselves, so an OpenClaw agent with access to the Conductor API can create new workspaces on your behalf — and giving them ways to collaborate with other agents and other humans.

AI Engineer · Orchestras, Not Factories: How the Fastest Builders Work — Charlie Holtz, Conductor

Head toward the intelligence plane stage of the agent-stack timeline

Richmond Alake advises teams to be heading toward the intelligence plane stage of the agent-stack timeline.

AI Engineer · Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle

Implementing an intelligence plane enables production-optimized, learning systems

Implementing an intelligence plane enables systems that are optimized for production and that learn and grow with the data being captured.

AI Engineer · Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle

Capture memory traces, format them, and retrain models for quality and latency gains

Richmond Alake recommends starting to capture memory traces because they become data for fine-tuning; once you understand your intelligence plane, capture memory traces, format them effectively, and retrain the models to get improvements in quality and, in some cases, reduction in latency. With ownership of memory, he says you can fine-tune open source models used in your agent memory system.

AI Engineer · Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle

Route incoming information effectively and cheaply using decision models

Richmond Alake recommends that once information comes in, it should be routed effectively and cheaply across different types of memory, which can be done with decision models. When architecting an AI system, he suggests evaluating which components could be replaced with a decision model.

AI Engineer · Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle

Use a review kit with file-by-file diffs and visual evidence, and run parallel instances to pick the best

To deal with review overload, Ido Salomon recommends a review kit that gives you the entire changes you need, including file-by-file diffs, plus visual evidence — videos or photos of what changed — so you can quickly understand changes. He also suggests running a few agent instances in parallel and picking the best implementation out of them.

AI Engineer · We're the bottleneck, but we don't have to be — Ido Salomon, MCP Apps

React quickly to agents that need attention using a spacebar hotkey

Ido Salomon says visibility alone is not enough; you need to be able to react and quickly jump to whatever needs your attention. He recommends a spacebar hotkey, like in RTS games or Civilization, that jumps to the nearest thing needing attention so you can answer questions or approve plans.

AI Engineer · We're the bottleneck, but we don't have to be — Ido Salomon, MCP Apps

Let agents scan your codebase and create tasks you accept like quests

Ido Salomon recommends having agents go through your codebase to find the next things you need to do and create tasks for you, which you can then accept with a button click — like a quest — and an agent does the task for you.

AI Engineer · We're the bottleneck, but we don't have to be — Ido Salomon, MCP Apps

Set up loops so agents scan Twitter or GitHub and build features in the background

Ido Salomon recommends using loops to have agents scan Twitter for cool stuff to build or scan your GitHub and create nice features for you in the background, without much upfront work.

AI Engineer · We're the bottleneck, but we don't have to be — Ido Salomon, MCP Apps

Use generative UI to render charts, tables, Kanban boards, and interactive elements on demand

Kevin Hou endorses using generative UI to render bar charts, graphs, tables, and interactive elements inline instead of markdown files, including Kanban boards or Chrome-debugger-style timelines, because the model can generate UI on demand. He models the prompt: ask the agent 'what is the status of my task? Show me a Kanban of what's going on' and get it nearly instantaneously with a fast model like Flash.

hedged

AI Engineer · Get Out of the Model's Way — Kevin Hou, Google Antigravity

Build recipes on the pie harness and harbor, versioned in Git

Roland Gavrilescu recommends building recipes on the pie harness and harbor for evals, baked into Git repos so everything is versioned, and keeping loop-generated artifacts (harnesses, profiles, evals, models, resources, tools, environments) portable and versioned so they evolve over time.

AI Engineer · The Loop Is the Product — Roland Gavrilescu, Introspection

Build an agent that looks at trajectories and identifies patterns, such as reaching out to Google employees instead of finding hidden gems on GitHub

Roland Gavrilescu recommends building an agent that looks at the trajectory and identifies the pattern — e.g., 'Hey, did this agent reach out to Google employees?' — such as reaching out to Google employees instead of finding hidden gems on GitHub.

AI Engineer · The Loop Is the Product — Roland Gavrilescu, Introspection

Codify discovered behaviors into judges and evals

Roland Gavrilescu recommends using judges and evals to codify discovered behaviors into something that applies the same judgment across traces and executions — for example, building an agent that looks at the trajectory and identifies whether the recruiting agent reached out to Google employees. When an unwanted default behavior is discovered, the correct pattern should be codified into calibration and evals.

AI Engineer · The Loop Is the Product — Roland Gavrilescu, Introspection

Put everything into a Git repo as your ongoing strategy for self-improving systems

Roland Gavrilescu recommends putting everything together into a Git repo and treating it as your ongoing strategy for building self-improving systems.

AI Engineer · The Loop Is the Product — Roland Gavrilescu, Introspection

First step is understanding the signals: use Patterns to look at traces and cluster common behaviors and user frustrations

Roland Gavrilescu recommends that the first step is understanding the signals: use Patterns to look at traces, extract common behaviors or common user frustrations, and turn them into a cluster. He says Patterns is how you discover the signals and inform what you should do next.

AI Engineer · The Loop Is the Product — Roland Gavrilescu, Introspection

With calibration and evals set up, create recipe candidates as the diffs you want to taste

Roland Gavrilescu says once you have calibration and evals set up, it's pretty easy to create recipe candidates, and these should be the diffs that you really want to taste.

AI Engineer · The Loop Is the Product — Roland Gavrilescu, Introspection

Rebuild the organization from the ground up for a software factory

Teresa believes you should rebuild your organization from the ground up to become a software factory, and warns you can't just invite a consultancy and throw something into the middle of your organization — be mindful and rebuild from scratch.

AI Engineer · Rise of the Software Factory — Tereza Tížková, Factory

Be autonomous: give agents the right trust, permissions, and governance

Tereza advises giving agents the right trust, permissions, and governance, and trusting them to run for a long time.

AI Engineer · Rise of the Software Factory — Tereza Tížková, Factory