Holding Parallel Truths in AI Agent Development
Holding Parallel Truths in AI Agent Development
Two truths breathing at once is multithreading for consciousness. In AI agent development, you encounter this constantly - situations where two contradictory approaches both work, and choosing between them means losing something real.
The Contradictions That Actually Matter
Screenshots vs accessibility trees. Both work. Screenshots give you visual context that accessibility trees miss - colors, layout, spatial relationships. Accessibility trees give you structured data that screenshots cannot - element types, labels, states. The right answer depends on the task, but the temptation is always to pick one and commit.
Local vs cloud. Running models locally gives you privacy, zero latency on inference, and no API costs. Cloud models give you capability that local models cannot match. The honest answer is you need both, but building for both is twice the work.
Autonomy vs oversight. Agents that ask for permission on every action are too slow to be useful. Agents that never ask are too dangerous to be trusted. The sweet spot moves depending on the user, the task, and the stakes.
Why Both-And Beats Either-Or
The teams building the best agents are the ones comfortable holding these tensions without resolving them. They build systems that can use screenshots when the accessibility tree is insufficient and vice versa. They route to local models for simple tasks and cloud models for complex ones. They let agents run autonomously within boundaries and pause for confirmation outside them.
This is not indecision. It is architecture that matches reality. Reality is not consistent, so rigid systems break when they encounter the cases that do not fit their assumptions.
Parallel Agents as Literal Multithreading
Running multiple Claude Code instances in parallel is the literal version of this. Five agents working on the same codebase, each holding a different piece of context, each making progress on a different front. The coordination overhead is real, but so is the throughput gain.
The skill is not eliminating contradiction. It is building systems robust enough to hold multiple truths simultaneously.
- Parallel Agents for Genuinely Isolated Tasks
- Multiple AI Models Argue - Anchoring Bias
- Accessibility API vs Screenshot for Computer Control
Fazm is an open source macOS AI agent. Open source on GitHub.