The Role of AI Agents

Agents are execution partners, not autonomous systems making product decisions. This is true whether you're a team of ten or a solo builder with a terminal.

What agents do well:

• Execute thin slices based on well-contextualized specs.
• Draft specs and hypotheses based on direction and context you provide.
• Run tests and surface drift early.
• Accelerate research (analyzing data, finding patterns, synthesizing information).

What humans do (and agents don't):

• Design experiments (what should we learn next?).
• Interpret evidence (what does this result actually mean?).
• Make judgment calls (double down, pivot, or abandon?).

For solo builders, agents are your team. They amplify your ability to build, but they don't replace your responsibility to learn. The vibe coder who lets the agent build whatever "feels right" without a hypothesis is just automating waste.

For teams, agents are shared tools, not owned by any one role. An engineer might direct an agent to build a thin slice. A PM might use an agent to draft a hypothesis. A designer might use one to prototype a validation experiment. The critical meta-skill for everyone: knowing when to let the agent just do it vs. when to carefully review. This judgment improves with practice and belongs to whoever is closest to the work.

Token efficiency matters

Agents aren't free. Every interaction costs tokens, and poorly contextualized work produces drift: the agent builds something off-target, you correct it, it rebuilds, you correct again. Each round burns tokens and time.

Good framing and clear slice specs are the best token optimization strategy. When the agent has strong context (the hypothesis, the success criteria, the technical constraints, the codebase patterns), it gets things right earlier and drifts less. Context engineering isn't just an intellectual principle. It has a direct cost impact.

Spec-Driven Development as the execution layer

If you practice Spec-Driven Development (SDD), it fits naturally inside LDD as the execution discipline for the Build phase. SDD provides the rigor to get clean, reliable output from agents. LDD provides the learning system that ensures the spec is worth building in the first place. Together they cover the full cycle: from hypothesis to validated decision, with disciplined agent execution in between.