How to burn tokens effectively.
From writing code to an agent swarm.
The higher you climb, the bigger the loop.
01A verified result. Less work carried between steps.
Manual handoffs
Code delegated.Handoffs still manual.
The patch came back. The human still ran the checks, copied the failure and requested another try.
Human
- Run the check
- Copy the failure
- Prompt again
Agent
- Write a patch
- Wait for feedback
- Write another patch
Every handoff brings the work back to you.
Practical experience
Delegate a bigger loop.
Same person. Different loops.
A bigger loop in software engineering does not automatically carry into making slides.
Software engineering
You know the system, the checks and what to delegate.
Making slides
You are learning the audience, visual standards and what to delegate.
A bigger loop is learned in each area.
04Level 0
Chat stops at a reply.
No agent loop. You carry the workflow.
Level 1
Two habits. Same missing loop.
Steer every edit
Super micromanagement.
Accept without checking
Vibe coding.
Level 2
The rising floor
The baseline: a verified task.
Software engineering baseline: delegate a complete task, then judge the evidence.
Level 2
The agent’s task loop
Let it finish the task.
Let the agent inspect the work, make the change, run the checks and repair failures inside the agreed boundary.
Sources: OpenAI · Agent loop
08You define done, permissions and the review.
Level 2
Connect the loop to the work.
Sources: Agent loop · MCP
Level 2
The worse gap:never seeing the possibility.
Level 2
The solution
Build the loop around the outcome.
Low 2.x loops around code. High 2.x follows the change through to its impact.
You hold scope, approvals and judgment. Agents carry authorized work and return evidence.
Scope feeds capability. Capability expands scope.
Level 3
Several agents.One shared goal.
Level 3
/goal · a strong evaluation
Give the goal a test of done.
Define an observable outcome and the checks that can show it was reached. A patch alone is not the result.
The test of done
Observable outcome + a check
Only the work and access needed for the agreed goal.
Level 3
/goal · checkpoints and scope
Each delivery earns the next step.
Deliver the agreed result, evaluate it and revise until it passes. Leave unrelated work outside the goal.
Only the work needed to meet the goal.
Level 3
Overlord v1 · fit and cognitive load
A swarm fits the workload. It adds a burden too.
Consider it for many concurrent tasks and constant interruptions. More contexts and decisions still demand judgment.
Most engineers will get more from a stronger task loop first.
Level 3
Overlord v1 · coordination
Many threads. One durable record.
Separate sessions do the work. Persistent task records hold objectives, dependencies, evidence and history.
The record connects work; the sessions do not share one model context.
Level 3
Overlord v1 · live demo
Overlord.
Live demo
- One bounded goal
- Separate task sessions
- Evidence back to you
Watch the brief, the handoff and the evidence returning.
Level 4
Overlord v2
The setup travels.
Level 4
Overlord v2 · inspecting its setup
Separate the method from the machine.
Each agent inspected its instructions to separate reusable methods from machine-specific context.
Level 4
Overlord v2 · illustrative instruction edit
It edits its owninstructions.
A shared method can keep a local form. Different machines can retain different behavior.
Shared method
Run the relevant checks before delivery.
Local instruction
Use the check commands in this machine’s runbook before delivery.
Same verification habit. Different local commands.
Instruction editing, not a change to model weights.
Level 4
Upgrade the shared base.Keep local intent.
Level 4
Overlord v2 · analogy
One base. Different local forms.
A biological analogy for shared instructions and retained differences across work environments.
| Biology | Overlord | |
|---|---|---|
| DNA | → | Shared base |
| Mutation | → | Self-edits |
| Environment | → | Machine + work |
| Strain | → | Local variant |
Engineering changes here are deliberate, reviewed and tested.
Level 4
Local adaptation
No task foundto use this yet.
Potential use case: work across different products or areas.
Level 5
Speculation · personal context
Imagine one mind that knows your world.
Assume a personal superintelligence with context across your life, work, preferences and connections.
- Life + work
- Thinking + preferences
- People + connections
Level 5
Speculation · automatic differentiation
Imagine copies that adapt themselves.
Each copy would change its code, tools and instructions to fit a different niche in your world.
Level 5
Speculation · coherent cooperation
Many specialized selves. One human direction.
Imagine those copies cooperating at vast scale and speed, guided by your goals and limits. Results return to you.
Your next experiment
Give one taska bigger loop.
Use your next experiment to move a complete piece of work, with its checks, inside a boundary you can hold.
Choose one task
A recurring piece of real work.
Define done
An observable result and its check.
Enable the work
Context, tools and explicit limits.
Review the evidence
Decide what the result earns next.
To give your agents a bigger loop, learn to hold a bigger one yourself.