RLabs Blog

Notas del laboratorio.

Foundations, implementations, patterns e industries. Lo que pensamos mientras construimos — sin filtros corporativos.

1 de junio de 2026
For 70 years the binding constraint was typing
1 de junio de 2026
Mixer Mode in 90 seconds, for the person at the console
1 de junio de 2026
The hat was a degenerate case, not a different mode
1 de junio de 2026
You suppressed three channels to ship one
1 de junio de 2026
Why I dropped the conductor and kept the mixer
1 de junio de 2026
Mixer Mode is not multitasking, and not coordination either
1 de junio de 2026
What I mean by Meta-Software (and what I don't)
1 de junio de 2026
The artifact shifted under your feet
1 de junio de 2026
What changes when MCP is the only way in
1 de junio de 2026
Security is where the seam would show, if there were one
1 de junio de 2026
The mixer does not rest by default
1 de junio de 2026
Self-awareness is the recursion stop
1 de junio de 2026
Human-in-the-loop is the setpoint, not the safety net
1 de junio de 2026
Software's Deming moment is already here
1 de junio de 2026
The METR slowdown is the predicted symptom, not the refutation
1 de junio de 2026
The pyramid is narrowing at the base
1 de junio de 2026
I wrote the framework precise enough to be wrong
1 de junio de 2026
What would actually refute this
1 de junio de 2026
If it's hard to articulate, that's the evidence — not the weakness
1 de junio de 2026
What I built when 'ask the LLM to respect hexagonal' kept failing
1 de junio de 2026
From prompt to primitive: why I stopped iterating on prompts
1 de junio de 2026
The 13 dimensions I use to score LLM-generated code
1 de junio de 2026
What it actually costs to embed quality contracts at every stage
1 de junio de 2026
Why I split code generation into 4 stages instead of one prompt
1 de junio de 2026
Building an MCP server so my IDE could do structural code review
1 de junio de 2026
What 19 MCP tools look like when you design them as a pipeline
1 de junio de 2026
Shipping one feature across six repos without losing my mind
1 de junio de 2026
Why I made the knowledge base headless
1 de junio de 2026
How I organized a super-agent by planes instead of layers
1 de junio de 2026
Evidence atoms: why every agent decision needs a causal chain
1 de junio de 2026
When the system writes its own architecture docs
1 de junio de 2026
One archetype, one artifact
1 de junio de 2026
Your cron jobs should write markdown: the automation-evidence loop
1 de junio de 2026
Atalaya + Akopia + AgentGuard: Meta-Software running on itself
1 de junio de 2026
MCP-readiness by default: every new service ships with a tool surface
1 de junio de 2026
The destructive flag that stopped an agent from wiping prod
1 de junio de 2026
Postgres RLS from day one, before the second tenant exists
1 de junio de 2026
Mixer Mode for CFOs — a thought experiment, not advice
1 de junio de 2026
Mixer Mode for SREs — a thought experiment, not advice
1 de junio de 2026
Mixer Mode for lawyers — a thought experiment, not advice
1 de junio de 2026
Mixer Mode in healthcare — a thought experiment, not clinical guidance
1 de junio de 2026
Mixer Mode for consultants — a thought experiment, not advice
1 de junio de 2026
Hexagonal by rubric, not by linter: a different way to enforce architecture
1 de junio de 2026
Mixer Mode for journalists — a thought experiment, not advice
1 de junio de 2026
Mixer Mode for translators — a thought experiment, not advice
1 de junio de 2026
Mixer Mode in the public sector — a thought experiment, not policy advice
1 de junio de 2026
Mixer Mode in financial back-office — a thought experiment, not advice
1 de junio de 2026
Industries where the framework does not reach
1 de junio de 2026
A 5-level maturity model for Mixer Mode (use it on yourself first)
1 de junio de 2026
A 4x3 audit to see where your Meta-Software layer leaks
1 de junio de 2026
From RPA scripts to Meta-Software: the 6-stage path I keep seeing
1 de junio de 2026
The 2x2 that tells you if your AI loop actually closes
1 de junio de 2026
An onboarding protocol for Mixer Mode (working hypothesis)
1 de junio de 2026
Evaluating Mixer Mode capacity in an interview (working hypothesis)
1 de junio de 2026
The mixer does not rest by default: practices I am testing for cognitive recovery
1 de junio de 2026
What 'a good task spec for an agent' looks like (a sub-articulated discipline)
1 de junio de 2026
I keep the diagnosis deflationary on purpose
1 de junio de 2026
When one search engine isn't enough: vector, BM25, and a coordinator
1 de junio de 2026
How do you know a practitioner actually crossed the transition?
1 de junio de 2026
Meta-Software is not SRE, DevOps, or Policy-as-Code rebranded. Here is why I think the difference matters
1 de junio de 2026
Audit your own stack: a guided self-check for the 4 Meta-Software categories
1 de junio de 2026
The channels are an example, not the count
1 de junio de 2026
The pillars only work together (and the example is incidental)
1 de junio de 2026
The four sub-categories of Meta-Software, and where each one breaks

For 70 years the binding constraint was typing

Mixer Mode in 90 seconds, for the person at the console

The hat was a degenerate case, not a different mode

You suppressed three channels to ship one

Why I dropped the conductor and kept the mixer

Mixer Mode is not multitasking, and not coordination either

What I mean by Meta-Software (and what I don't)

The artifact shifted under your feet

What changes when MCP is the only way in

Security is where the seam would show, if there were one

The mixer does not rest by default

Self-awareness is the recursion stop

Human-in-the-loop is the setpoint, not the safety net

Software's Deming moment is already here

The METR slowdown is the predicted symptom, not the refutation

The pyramid is narrowing at the base

I wrote the framework precise enough to be wrong

What would actually refute this

If it's hard to articulate, that's the evidence — not the weakness

What I built when 'ask the LLM to respect hexagonal' kept failing

From prompt to primitive: why I stopped iterating on prompts

The 13 dimensions I use to score LLM-generated code

What it actually costs to embed quality contracts at every stage

Why I split code generation into 4 stages instead of one prompt

Building an MCP server so my IDE could do structural code review

What 19 MCP tools look like when you design them as a pipeline

Shipping one feature across six repos without losing my mind

Why I made the knowledge base headless

How I organized a super-agent by planes instead of layers

Evidence atoms: why every agent decision needs a causal chain

When the system writes its own architecture docs

One archetype, one artifact

Your cron jobs should write markdown: the automation-evidence loop

Atalaya + Akopia + AgentGuard: Meta-Software running on itself

MCP-readiness by default: every new service ships with a tool surface

The destructive flag that stopped an agent from wiping prod

Postgres RLS from day one, before the second tenant exists

Mixer Mode for CFOs — a thought experiment, not advice

Mixer Mode for SREs — a thought experiment, not advice

Mixer Mode for lawyers — a thought experiment, not advice

Mixer Mode in healthcare — a thought experiment, not clinical guidance

Mixer Mode for consultants — a thought experiment, not advice

Hexagonal by rubric, not by linter: a different way to enforce architecture

Mixer Mode for journalists — a thought experiment, not advice

Mixer Mode for translators — a thought experiment, not advice

Mixer Mode in the public sector — a thought experiment, not policy advice

Mixer Mode in financial back-office — a thought experiment, not advice

Industries where the framework does not reach

A 5-level maturity model for Mixer Mode (use it on yourself first)

A 4x3 audit to see where your Meta-Software layer leaks

From RPA scripts to Meta-Software: the 6-stage path I keep seeing

The 2x2 that tells you if your AI loop actually closes

An onboarding protocol for Mixer Mode (working hypothesis)

Evaluating Mixer Mode capacity in an interview (working hypothesis)

The mixer does not rest by default: practices I am testing for cognitive recovery

What 'a good task spec for an agent' looks like (a sub-articulated discipline)

I keep the diagnosis deflationary on purpose

When one search engine isn't enough: vector, BM25, and a coordinator

How do you know a practitioner actually crossed the transition?

Meta-Software is not SRE, DevOps, or Policy-as-Code rebranded. Here is why I think the difference matters

Audit your own stack: a guided self-check for the 4 Meta-Software categories

The channels are an example, not the count

The pillars only work together (and the example is incidental)

The four sub-categories of Meta-Software, and where each one breaks