The Greenhouse and the Lens: Two Modes of Agentic AI Work

4 hours ago

Agentic AI can be understood through two contrasting modes: greenhouse (diffuse growth) and lens (focused ignition), both powered by 'sunlight' or AI capabilities.
Greenhouse mode is ideal for exploration when goals are unclear, allowing many tools or ideas to grow simultaneously, with low cost enabling experimentation and post-hoc evaluation.
Lens mode involves concentrating effort on a well-defined target, discarding non-essential proposals, and is useful for precise execution when 'done' is clear.
Neither mode is inherently superior; they serve different purposes, and mismatching mode to task causes frustration with agentic AI.
Advanced expertise includes recognizing which mode a task demands, staying in that mode, and correcting drift (e.g., accidental feature creep or over-polishing during exploration).
Domain expertise acts as an 'oracle' in both modes: sifting promising results in greenhouse and rejecting plausible-but-wrong outputs in lens mode, while also detecting mode drift.

The author initially doubted AI's coding abilities, treating it like Stack Overflow and manually reviewing code to maintain personal style.
Claude Code demonstrated AI's ability to handle maintenance tasks and reduce the need to learn specific frameworks like Laravel.
Newer AI models (e.g., Fable, GPT 5.6-Sol) now outperform the author in 80% of coding tasks, leading to acceptance and excitement about productivity gains.

Schema harness enables frontier models to achieve ~99% performance on ARC-AGI-3 Public by treating latent world representations as executable programs, allowing interpretable, verifiable, and searchable modeling.
ARC-AGI-3 challenges agents to infer game rules, objects, and goals from raw 64x64 color grids without explicit instructions, using the Relative Human Action Efficiency (RHAE) metric for evaluation.
The approach combines state grounding (identifying objects and variables from observations) and mechanism discovery (learning transition rules), solved jointly within an editable program that mirrors physicist-like reasoning.

Invest in upfront alignment by making implicit assumptions explicit in the initial prompt to avoid unexpected AI outputs.
Restart with better context rather than steering AI when initial outputs are far off; avoid compounding errors through path dependence.
Equip AI agents with the same tools you use (CLI, MCP, browser control) to handle setup, testing, and instrumentation tasks.

For many agent tasks, the best 'tool' is a small code-execution surface, allowing models to write ordinary code against typed capabilities and enforce security below the model.
Traditional agent toolboxes are growing with many tools, but models already excel at code, which is a more familiar and efficient interface, reducing context usage and tool calls.
Code mode compresses context (e.g., 88% less), reduces token consumption from tool schemas, and allows deterministic logic between calls, such as joins, filters, and error handling.

Thinking Machines aims to develop AI that enhances human will and judgment, ensuring AI serves human purposes shaped by continuous feedback and learning.
Current AI often lacks customization from users; the vision is distributed, diverse AI that learns alongside people, not centralized, frozen models.
Key technical directions include training strong, multimodal models; tools for customization (e.g., training model weights); interactive interfaces for live collaboration; and publishing research for broader understanding.

The author initially doubted AI's coding abilities, treating it like Stack Overflow and manually reviewing code to maintain personal style.
Claude Code demonstrated AI's ability to handle maintenance tasks and reduce the need to learn specific frameworks like Laravel.
Newer AI models (e.g., Fable, GPT 5.6-Sol) now outperform the author in 80% of coding tasks, leading to acceptance and excitement about productivity gains.

Schema harness enables frontier models to achieve ~99% performance on ARC-AGI-3 Public by treating latent world representations as executable programs, allowing interpretable, verifiable, and searchable modeling.
ARC-AGI-3 challenges agents to infer game rules, objects, and goals from raw 64x64 color grids without explicit instructions, using the Relative Human Action Efficiency (RHAE) metric for evaluation.
The approach combines state grounding (identifying objects and variables from observations) and mechanism discovery (learning transition rules), solved jointly within an editable program that mirrors physicist-like reasoning.

Invest in upfront alignment by making implicit assumptions explicit in the initial prompt to avoid unexpected AI outputs.
Restart with better context rather than steering AI when initial outputs are far off; avoid compounding errors through path dependence.
Equip AI agents with the same tools you use (CLI, MCP, browser control) to handle setup, testing, and instrumentation tasks.

For many agent tasks, the best 'tool' is a small code-execution surface, allowing models to write ordinary code against typed capabilities and enforce security below the model.
Traditional agent toolboxes are growing with many tools, but models already excel at code, which is a more familiar and efficient interface, reducing context usage and tool calls.
Code mode compresses context (e.g., 88% less), reduces token consumption from tool schemas, and allows deterministic logic between calls, such as joins, filters, and error handling.

Thinking Machines aims to develop AI that enhances human will and judgment, ensuring AI serves human purposes shaped by continuous feedback and learning.
Current AI often lacks customization from users; the vision is distributed, diverse AI that learns alongside people, not centralized, frozen models.
Key technical directions include training strong, multimodal models; tools for customization (e.g., training model weights); interactive interfaces for live collaboration; and publishing research for broader understanding.

Hasty Briefsbeta