How Adam works
with AI agents

Adam writes one short message, and his agents plan, build and prove the work.

Adam is a developer and educator. He runs 10 AI agents. Many people ask how he works with them. This page shows how, and what you can copy.

  1. 1AdamOne short line, often by voice.
  2. 2Grok Bot agentA chat agent that runs for a long time, such as Tony or Johnny. It writes the task, not the code.
  3. 3CoordinatorA job that Limen starts. It owns the spec (the written plan), splits it into tickets and runs the workers.
  4. 4WorkersJobs that build the parts. The hardest tasks go to a group of models from three providers.
  5. 5EvidenceTests, screenshots or a live run show that it works.
  6. 6AdamGets a Ready message with steps to try, or sends a correction.

Limen is Adam's job tool. It starts, steers and merges the jobs. It is the main part of his harness, the software around the models. Based on 4,547 messages from Adam to his agents, Sep 1 to Oct 20, 2026.

From one message to verified work

The full path of one task. Press Play, or click a step.

One message starts work that Adam does not write

How Adam thinks about agents

Six ideas sit under his 14 moves.

01Work is done only when it is proven

Merged work is not done work. Work is done when evidence shows that it works. The lab is a test version of the app that never ships.

No evidenceEvidence

In your first sentence, answer me: was the test done in the lab with a real API key?Adam to Tony, Sep 12
8.6%of Adam's messages ask for verification: about one in 12. He asks for proof again and again.
Why it works. A report can pass while the product fails. Evidence is the only status that Adam does not have to trust.

02Give tasks to coordinators, not to workers

Adam does not give tasks to workers. He gives a task to a Grok Bot agent. The Grok Bot agent starts a coordinator. The coordinator owns the spec and runs the workers.

Why do you start workers and not coordinators? We always start coordinators that manage workers. Workers do not have enough knowledge to manage the specs.Adam to Tony, Sep 8
Why it works. A worker has a task but no spec. A coordinator owns the spec, splits the work and checks the result. Adam talks to a few coordinators, not to many workers.

03Write the global context once, then use simple words

The vision and the style guide describe the whole product. Adam writes them once and updates them from time to time. Because of them, a few simple words from Adam are enough.

The vision · global

What the product is and where it goes. Written once. Agents read it before they choose work.

The style guide · global

The rules for code and text in the whole product. Written once. The harness adds it to every call.

Semantic anchors

Short, heavy words that the model already knows well, such as "primitive" or "authority". One word pulls in a whole concept.

The board

The board mixes the global context with the local context of each feature: the decisions, the work now, and what is proven. It also shows each agent what the other agents do. Click a part.

When Adam talks about one feature spec, the agent reads it in three ways: through what Adam says, through what is already in the project, and through the global context.

Why it works. Adam explains the product one time, in files. After that, each message only needs to name the change. The agent adds the rest from the files.

04Discuss the feature, then let the agents plan the tasks

Adam discusses a feature, or a set of features, with his agents. The agents then split it into tasks, set the order and the priorities, and keep the board.

One possible path. The group decides its own phases.

← Questions go back to Adam. His answers+ add− drop½ reduceslices.

Two real features

In one project, Sep 1 to Oct 20: 62 tickets and 139 jobs.

Why it works. Adam spends his time on product decisions. The agents do the planning: the tickets, the order, the priorities and the board.

05Short messages work because the context is shared

Adam's messages can be very short. The harness adds the context around each message.

I need groups fixed, not removed.6 words
23.9%of Adam's messages have five words or fewer: about one in four. A short message is normal.

What the harness adds

Layers of context around the message, before the agent reads it. Click a layer.

Why it works. The harness puts the rules, Adam's earlier words and the evidence around each short message. Adam does not have to write them again.

06Agents decide by the product vision, not by taste

The agent makes the decision. It uses the vision and the work around it first. It uses its own taste last.

Start the coordinators for sign-in, trial, purchase and onboarding. Make reasonable choices. Use the vision, and remember that we prefer simple, primitive-based, general solutions.Adam to Tony, Sep 9
Why it works. Each coordinator starts with one rule: engineering time is cheap and Adam's attention is scarce. So the coordinator decides and verifies. Deploys still need Adam's approval.

07Steer by how models and harnesses work

Adam knows how models and harnesses handle context. He helps each agent find the right context. He does not give very specific instructions.

I steer agents toward exploration first, because a model answers based on what is already in its context.Adam, voice note, Oct 20
Why it works. The agent builds its own context from the right sources. Adam gives the direction, not every step.

08Let one agent audit and fix another

When an agent works badly, Adam does not correct the reply. A second agent audits the first one and fixes the cause: the prompt, the handoff, the rule or the tool.

Auditor Targetreplies reads the output writes the fix
Why it works. A fix in the prompt, the rule or the tool changes all future replies. A correction changes only one reply.

09Use models from three providers on difficult tasks

For the most difficult tasks, Adam asks for a group: one agent each from xAI (Grok), OpenAI (Sol) and Anthropic (Opus). The group plans its own process and returns one result.

The group picks:

Example only. The group chooses its own phases for each task.

Why it works. Different providers give different views of the same problem. The group compares them and returns one result. Groups are only for the hardest tasks, such as design decisions and specs. One model does the normal build work.

10Make all agents write plain technical English

All agents write plain technical English: about 80% of ASD-STE100, a controlled language for technical manuals. They write for an operator who does not read code.

Short sentencesOne idea in each sentence.
Active voice"The coordinator merges it." Not "It was merged."
One word, one thingUse the same word for the same thing.
No fillerNo slang and no idioms.
No internal namesNo ids, hashes or paths unless Adam asks.
Do not use the phrase "on tip". Say that the work is on main and ready to test.Adam to Tony, Sep 12
Why it works. Adam does not see the board or the files. The reply is the only thing he sees.

11Dictate long messages, type short ones

Adam dictates when he tests the app, when he does other work and when he drives. He types short tasks and dictates long ones.

~24% of messages
~51% of words

A quarter of the messages carries half of the words. The dictated messages are the long ones.

Typed · 7 words
Check on them, and steer if needed.
To Tony · Sep 5
Dictated · 161 words
The error, and basically all the UI here, is really poor, and this error message is very poor too. Also, when I'm in settings, click Files and then go back, I think I should land in settings again, not in the chat, because this way navigation is broken. "Reconnect" in settings: I'm not sure where to reconnect. It asks me to reconnect a folder in settings, but I'm already in settings. What's going on here? I'm not even sure how to manage it. I'm speaking out loud, so can you read this? We need to spin a coordinator to work out the issues with folders. The experience isn't smooth. It has to understand the navigation, the settings, all the panels, consistency and so on. Spin Claude for this. Something is really off here, and I need to find out exactly what. The result has to be a specification file built on common sense and reasoning about the logic we have here.
To Tony · Sep 13
Break down what I just said. Before you do anything, let me confirm that you understood me. Be brief.Adam to Tony, Sep 11, at the end of a dictated message
Why it works. Dictation can contain errors. A short read-back finds a wrong word before the work starts.

12Decide from a visual page, not a report

Adam discusses a topic with his general assistant. Then he asks for an HTML page and makes his decision from it.

  • Black and white
  • Visuals, not textGraphs, diagrams and toggles.
  • A live preview of each stateEvery state, not only the normal one.
  • A plain explanation of each behaviorOne or two short sentences.
  • Design options to choose from
I need an HTML page in my browser with a clean graph of the architecture. I need to click the parts to see what we have and what is missing.Adam to Tony, Sep 10

Each state

Ready
Waiting
Blocked
Parked

Design options

Why it works. Adam can judge a picture in seconds. A written report takes minutes to read.

13Automate the small tasks around each job

Limen does the small tasks around each job. The agents and Adam do not have to do them. Click a step.

Some coordinators stopped or are stuck. Look at them and stop them. But first, record where they stopped, so that we can restart them.Adam to Tony, Sep 9
Why it works. Each task that runs alone is one less thing for Adam or an agent to remember.

14Engineer the system so that agents can move fast

When agents do the building, the structure of the system sets the speed. Adam designs the structure. The agents write the code.

  1. Start from a vision that names the parts.
  2. Give each job one part. Remove the copies.
  3. Give agents a few primitives with one shape: find, spec, call.
  4. Agents and the UI use the same requests.
  5. Keep a lab that never ships, on the same paths as the real app.
  6. Map the system from the vision down to line counts.
  7. Tidy up toward the vision. Do not rewrite.
Why it works. An agent finds the correct part and does not write a new copy. It proves its work in the lab before it reports.

Example: Alice, one of Adam's apps

240,949lines of source code. Agents wrote them.
188,582lines ship in the Mac app.
44,300lines are in the lab and never ship.

Almost one line in five is lab code. The agents need a place to prove their work.

Try this yourself

Patterns from Adam's messages. Fill in the brackets and use them with your own agents.