Blog

An evaluation rubric is a theory of good work

Before an AI judge scores an agent, ask the person who knows the task which outcomes count and which errors disqualify the result.

· 4 min read

3 sections

3 sections

A useful map leaves things out

A diagram helps when it preserves the relationships needed for a task. Name what it distorts and offer a companion view when another task needs those details.

· 4 min read

Station A → Transfer → Station B

Station A → Transfer → Station B

Why a hundred nearly perfect steps still fail

A 99% reliable step repeated a hundred times succeeds only 36.6% of the time under a simple model. Here is what that means for long agent workflows.

· 4 min read

3 sections

3 sections

What is each approval protecting?

Every approval can guard a real risk or add waiting. Map the decision path before removing or adding a gate.

· 4 min read

PREFLIGHT → ACTIVATION

PREFLIGHT → ACTIVATION

Ask the source before copying the data

A question may need a fresh read from the source, not a second database. Choose remote access or sync by what the next system must do.

· 4 min read

Source of truth → Scoped query → Consumer

Source of truth → Scoped query → Consumer

The person who knows the exceptions should shape the automation

A workflow specification loses the cases that make work hard. Let the task expert define those cases while code enforces data and permission rules.

· 4 min read

Task expert → Reviewed rule → Execution

Task expert → Reviewed rule → Execution

Cash runway is a race between revenue and the bank balance

A worked cash runway example shows why current losses and survival are different questions, and how fragile a growth projection can be.

· 5 min read

3 sections

3 sections

Put changing facts where they can be corrected

A product catalog or API contract changes. If it lives in an agent prompt, every correction becomes a prompt edit and every call pays to reread it.

· 4 min read

3 sections

3 sections

Check everything cheaply; investigate the few that need it

A broad first pass can keep costly judgment focused on uncertain cases. The design depends on what the cheap check misses, not just what it saves.

· 4 min read

FIRST → ONLY IF NEEDED

FIRST → ONLY IF NEEDED

Make the constraint visible before choosing the work

A visible limit turns a backlog into a choice. Show what each new request displaces before calling it a priority.

· 4 min read

3 sections

3 sections