THINKING

Notes on engineering leadership, AI, and the future of how we build software.

Last updated: July 2026

BUILDING TOWARD AGENTIC DEVELOPMENT :

From force multiplier to collaborative force, and what still requires human judgment

After introducing significant AI-assisted development at vivimetrix, I began thinking more seriously about what it would take to move from AI as a productivity tool to AI as a genuine participant in the software delivery process. I have watched teams both underuse and over-trust these systems. I have also seen the failure modes, from poor deployments to more serious production incidents, when guardrails and clear ownership are missing.

The core issue is not that current AI systems are incapable of writing code, generating tests, or proposing designs. In many cases they are surprisingly capable. The deeper limitation is the absence of judgment that comes from lived experience. Humans carry an intuitive sense of consequence built over years of seeing what happens when things go wrong. Rules, prompts, and context windows are useful, but they are not a complete substitute for that experience.

WHERE AI CAN HELP TODAY :

Looking at a typical software delivery lifecycle, the opportunities and limits are uneven:

Product direction and problem framing: Still heavily human. AI can assist with research and exploration, but novel judgment and prioritization remain difficult to delegate.
Requirements and nuance: High human involvement required. Context and trade-offs are often subtle.
Implementation: This is currently the strongest area. AI can take clear requirements and produce solid first versions of UI, APIs, data models, and boilerplate at high speed.
Code review: Useful for consistency, patterns, and basic quality. Still benefits from human oversight on architectural and contextual issues.
Testing: AI can generate substantial test coverage. Humans remain important for deciding what actually matters to test.
Deployment and production changes: Highest need for caution. This is where missing judgment carries the greatest cost.

TOWARD A MORE AGENTIC MODEL :

A practical path forward is not to wait for a single general agent that can do everything. It is to build specialized agents that become highly competent at individual stages of the lifecycle, then gradually increase their scope as we improve evaluation, guardrails, and feedback loops.

In this model, AI systems are treated more like capable junior or mid-level engineers who need clear constraints, strong review practices, and well-defined ownership. The goal is not full autonomy for its own sake. The goal is higher leverage while preserving accountability.

Metrics matter here. We already measure delivery performance in various ways. We will need equivalent ways to measure how well these systems understand requirements, produce reliable changes, and respond to failure. Those signals should feed back into the systems themselves.

WHAT REMAINS HUMAN :

Even as capability improves, certain responsibilities are likely to stay with people for a long time: setting direction, making value judgments, absorbing accountability, and deciding when the risk of a change is acceptable. The most effective teams will be those that get the division of labor right, using AI for speed and coverage while keeping humans firmly in the loop where judgment and ownership matter most.

I am interested in practical approaches that move us toward higher-leverage engineering organizations without pretending that the hard parts of software leadership have been solved.

THINKING

Notes on engineering leadership, AI, and the future of how we build software.

Last updated: July 2026

BUILDING TOWARD AGENTIC DEVELOPMENT :

From force multiplier to collaborative force, and what still requires human judgment

After introducing significant AI-assisted development at vivimetrix, I began thinking more seriously about what it would take to move from AI as a productivity tool to AI as a genuine participant in the software delivery process. I have watched teams both underuse and over-trust these systems. I have also seen the failure modes, from poor deployments to more serious production incidents, when guardrails and clear ownership are missing.

The core issue is not that current AI systems are incapable of writing code, generating tests, or proposing designs. In many cases they are surprisingly capable. The deeper limitation is the absence of judgment that comes from lived experience. Humans carry an intuitive sense of consequence built over years of seeing what happens when things go wrong. Rules, prompts, and context windows are useful, but they are not a complete substitute for that experience.

WHERE AI CAN HELP TODAY :

Looking at a typical software delivery lifecycle, the opportunities and limits are uneven:

Product direction and problem framing: Still heavily human. AI can assist with research and exploration, but novel judgment and prioritization remain difficult to delegate.
Requirements and nuance: High human involvement required. Context and trade-offs are often subtle.
Implementation: This is currently the strongest area. AI can take clear requirements and produce solid first versions of UI, APIs, data models, and boilerplate at high speed.
Code review: Useful for consistency, patterns, and basic quality. Still benefits from human oversight on architectural and contextual issues.
Testing: AI can generate substantial test coverage. Humans remain important for deciding what actually matters to test.
Deployment and production changes: Highest need for caution. This is where missing judgment carries the greatest cost.

TOWARD A MORE AGENTIC MODEL :

A practical path forward is not to wait for a single general agent that can do everything. It is to build specialized agents that become highly competent at individual stages of the lifecycle, then gradually increase their scope as we improve evaluation, guardrails, and feedback loops.

In this model, AI systems are treated more like capable junior or mid-level engineers who need clear constraints, strong review practices, and well-defined ownership. The goal is not full autonomy for its own sake. The goal is higher leverage while preserving accountability.

Metrics matter here. We already measure delivery performance in various ways. We will need equivalent ways to measure how well these systems understand requirements, produce reliable changes, and respond to failure. Those signals should feed back into the systems themselves.

WHAT REMAINS HUMAN :

Even as capability improves, certain responsibilities are likely to stay with people for a long time: setting direction, making value judgments, absorbing accountability, and deciding when the risk of a change is acceptable. The most effective teams will be those that get the division of labor right, using AI for speed and coverage while keeping humans firmly in the loop where judgment and ownership matter most.

I am interested in practical approaches that move us toward higher-leverage engineering organizations without pretending that the hard parts of software leadership have been solved.

THINKING

Notes on engineering leadership, AI, and the future of how we build software.

Last updated: July 2026

BUILDING TOWARD AGENTIC DEVELOPMENT :

From force multiplier to collaborative force, and what still requires human judgment

After introducing significant AI-assisted development at vivimetrix, I began thinking more seriously about what it would take to move from AI as a productivity tool to AI as a genuine participant in the software delivery process. I have watched teams both underuse and over-trust these systems. I have also seen the failure modes, from poor deployments to more serious production incidents, when guardrails and clear ownership are missing.

The core issue is not that current AI systems are incapable of writing code, generating tests, or proposing designs. In many cases they are surprisingly capable. The deeper limitation is the absence of judgment that comes from lived experience. Humans carry an intuitive sense of consequence built over years of seeing what happens when things go wrong. Rules, prompts, and context windows are useful, but they are not a complete substitute for that experience.

WHERE AI CAN HELP TODAY :

Looking at a typical software delivery lifecycle, the opportunities and limits are uneven:

Product direction and problem framing: Still heavily human. AI can assist with research and exploration, but novel judgment and prioritization remain difficult to delegate.
Requirements and nuance: High human involvement required. Context and trade-offs are often subtle.
Implementation: This is currently the strongest area. AI can take clear requirements and produce solid first versions of UI, APIs, data models, and boilerplate at high speed.
Code review: Useful for consistency, patterns, and basic quality. Still benefits from human oversight on architectural and contextual issues.
Testing: AI can generate substantial test coverage. Humans remain important for deciding what actually matters to test.
Deployment and production changes: Highest need for caution. This is where missing judgment carries the greatest cost.

TOWARD A MORE AGENTIC MODEL :

A practical path forward is not to wait for a single general agent that can do everything. It is to build specialized agents that become highly competent at individual stages of the lifecycle, then gradually increase their scope as we improve evaluation, guardrails, and feedback loops.

In this model, AI systems are treated more like capable junior or mid-level engineers who need clear constraints, strong review practices, and well-defined ownership. The goal is not full autonomy for its own sake. The goal is higher leverage while preserving accountability.

Metrics matter here. We already measure delivery performance in various ways. We will need equivalent ways to measure how well these systems understand requirements, produce reliable changes, and respond to failure. Those signals should feed back into the systems themselves.

WHAT REMAINS HUMAN :

Even as capability improves, certain responsibilities are likely to stay with people for a long time: setting direction, making value judgments, absorbing accountability, and deciding when the risk of a change is acceptable. The most effective teams will be those that get the division of labor right, using AI for speed and coverage while keeping humans firmly in the loop where judgment and ownership matter most.

I am interested in practical approaches that move us toward higher-leverage engineering organizations without pretending that the hard parts of software leadership have been solved.