The Infrastructure Gap

Coding agents have achieved high levels of autonomy not just because of better models, but because the software engineering ecosystem provides a ready-made infrastructure for agents. Knowledge work, by contrast, lacks the primitives necessary for safe, autonomous operation. To bridge this gap, we must replicate six core primitives that coding already enjoys.

The Six Primitives for Autonomous Agents

1. Centralization and History

Coding agents benefit from a single source of truth (the codebase) and a permanent record of changes (Git). Knowledge work is fragmented across disparate platforms (Slack, Salesforce, Notion, Gmail), forcing agents to "stitch" data together. By centralizing these connections, agents gain a baseline of context. Furthermore, logging every agent action creates a "memory" that allows the agent to learn from past successes and failures, rather than starting from a blank state every time.

2. Context and Verification

Agents need two types of context: the structural map of how an organization functions and the stylistic preferences of the user. In coding, linters and type systems provide this. In knowledge work, this context is often implicit. Verification is equally critical; coding agents use unit tests and compilers to catch errors before they hit production. Knowledge work agents require similar "sandboxes" to mock real-world actions, allowing for human review before the agent executes irreversible tasks.

3. Governance and Reversibility

Governance must be enforced by external walls, not just internal prompts, which are fragile and can be "compacted away" by the model. Deterministic control over access (e.g., read-only vs. write permissions) and natural language policies (e.g., "never delete more than 10 emails") provide safety. Reversibility is the final, hardest hurdle. In code, Git allows for easy reverts. In knowledge work, many actions (like sending an email or wire transfer) are permanent. Where true undo is impossible, the infrastructure must force the agent to operate in a sandbox environment first, shifting the human review from 'post-action' to 'pre-action'.

Shifting the Bottleneck

The bottleneck for AI development has shifted away from the models themselves. The same models capable of writing code are currently "working blind" in knowledge work due to a lack of surrounding infrastructure. By building the necessary scaffolding—centralization, logging, sandboxing, and governance—we can enable knowledge work agents to achieve the same level of autonomy as their software engineering counterparts.