Blog
Notes on building AI agents that hold up
Field notes from building aibuddy in the open — the runtime, context engineering, and the coordination problems a bigger model doesn't solve for you.
Models Are Not the End: Agents and the Next Organizational Shift
The agent ecosystem is still waiting for an entry point that dramatically lowers the barrier to use. Models will absorb simple scaffolding, but not complex work—and the next advantage will come from task loops and intelligence flywheels.
Why AI agents need a harness, not just a better model
Better models raise the ceiling, but reliable agents come from the system around the model: context, tools, constraints, verification, correction, and observable loops.
AI-native companies reorganize context
AI-native transformation is not about adding AI to old workflows. It is about reorganizing company information so agents receive the right context before each decision and return evidence for the next one.
AI Jobs Transition (1): Capability is not displacement
How much work AI can perform tells us where change may begin, not whether a job will disappear, reorganize, or grow. We need a better framework than another automation-risk ranking.
AI Jobs Transition (2): The work that still needs a person
A job can be highly exposed to AI and still require a person for accountability, trust, or physical execution. That protects the role from full automation, but not from redesign.
AI Jobs Transition (3): Cheaper work can create more work
AI reduces the labor needed for each unit of output, but lower costs can also unlock new customers and new demand. Employment depends on which force wins.
AI Jobs Transition (4): The capability gap is an operating problem
Models can already affect far more work than organizations actually delegate to them. Closing that gap requires context, permissions, evaluation, and workflow redesign—not more prompt tips.
AI Jobs Transition (5): Four paths, not one future
The labor market is not moving toward one AI outcome. Four transition paths require different company decisions, worker strategies, and policy responses.
AI Jobs Transition (6): Turn the framework into an operating plan
The useful output of AI labor research is not a prediction. It is a repeatable way to choose what to automate, what to redesign, where to grow, and what to monitor.
Why enterprises need AI
Much of enterprise work depends on language, unstructured information, and professional judgment that traditional software cannot economically encode. Foundation models are turning that work into a general, scalable software capability.
Harness Engineering (1): When engineers start designing the agent's environment
When code generation is no longer scarce, engineering shifts from writing every implementation to designing environments, expressing intent, and building feedback loops.
Harness Engineering (2): Give the agent a map of the repository
A giant AGENTS.md does not solve context. Agents need a short entry point, layered knowledge, and a system of record that stays aligned with the code.
Harness Engineering (3): Make the running system legible to the agent
An agent that can edit code still cannot verify the product. Autonomous engineering requires applications, browsers, logs, metrics, and acceptance criteria that agents can inspect directly.
Harness Engineering (4): Turn engineering rules into executable constraints
Agents do not follow a rule forever because they read it once. A reliable harness turns architecture, quality requirements, and engineering taste into checks the system can enforce.
Harness Engineering (5): When agent output exceeds human attention
Once agents increase code throughput, one-by-one human review becomes the next bottleneck. Checks, review, merge policy, and recovery must be configured by risk.
Harness Engineering (6): Agents copy technical debt too
Agents learn from the patterns already in the repository, including duplication, drift, and temporary patches. High-throughput systems need continuous quality garbage collection.
AI Builders: From Using AI to Shipping Systems
AI builders do more than use tools well. They put AI into real settings and turn it into systems that can be delivered, verified, and improved.
Ideas are cheap. Execution is expensive.
In the AI era, many directions become obvious. The real advantage is not having the idea first, but defining the problem, filtering noise, building a verifiable execution system, and carrying long-horizon work to completion.
AI-Native Startup (1): What Actually Gets Rebooted
AI does not make startups easy. It compresses the loop from idea to evidence to product to feedback.
AI-Native Startup (2): Existing Company Data Is the First Asset
Before an AI-native company asks AI to write more code, it has to turn customer, product, sales, and ops data into reusable context.
AI-Native Startup (3): The Founder Becomes an Orchestrator
AI does not make the founder's job lighter. It moves the work from direct execution to system design, judgment, and orchestration.
AI-Native Startup (4): In the Idea Stage, Building Is Not Validation
AI makes building dangerously easy, so the Idea stage has to prove the problem is real, specific, frequent, and worth solving.
AI-Native Startup (5): In the MVP Stage, Boundaries Matter More Than Code
The first artifact of an AI-native MVP should not be code. It should be scope, architecture, metrics, and context.
AI-Native Startup (6): Launch Is Where Product Becomes Company
Launch is not the announcement. It is the stage where founder improvisation becomes a repeatable operating system.
AI-Native Startup (7): The Real Moat for AI Products
The moat is not the model. It is domain knowledge, user behavior data, workflow embeddedness, and the operating loop that compounds them.
AI-Native Startup (8): Same Founder Job, New Rules
AI does not replace founder judgment. It makes judgment, context, and learning speed more central than ever.
From using AI to becoming an AI-native team
AI-native teams are not defined by how many people use chat. They are defined by whether agents can participate in real work with shared context, scoped authority, verification, and accountable humans.
Build your company as an intelligence layer, not an org chart
AI shouldn't be a tool your company uses — it should be the operating system your company runs on, and that rewrites the org chart, not just the output.
How to build an AI-native services company
Some of the biggest companies of the next decade won't sell software — they'll be law firms, insurers, and tax practices rebuilt from scratch with AI doing most of the work.
How to pick a startup idea
The perfect idea doesn't exist in the abstract — the only way to find what works is to pick one, burn the other boats, and go deep enough to run your customer's business.
How to get your first ten customers
Your first ten customers almost never come from a tool — they come from your network, showing up in person, and a willingness to do the things that don't scale.
Context engineering is the bottleneck in long agent runs
Long agent runs fail when the model sees the wrong slice of state. The hard part is not stuffing the window; it is managing context as working memory across turns, tools, compaction, memory, and cache.
No posts in this category yet.