The hype around autonomous companies often fixates on selecting the right large language model, but Tej Pandya, founder of GrowEasy.ai, argues that the real differentiator is control-plane design. In a recent post on DEV.to, Pandya outlines a lane-based architecture for a hypothetical small service firm handling lead intake, emphasizing that autonomy is enforced through code rather than prompt engineering. The framework splits every operational task into one of three distinct lanes: Workflow, Agent, or Human, ensuring predictable behavior while allowing adaptive reasoning where necessary.

Enforcing Lanes in Code

The core mechanism involves a Python enumeration that strictly categorizes tasks. Workflow lanes handle fixed steps with no model decisions, such as capturing leads via CRM creation or sending template emails. Agent lanes allow the model to pick steps within defined limits, like drafting replies using specific CRM and email tools. Human lanes reserve decisions for people, particularly for high-stakes actions like quoting prices. Pandya notes that Anthropic’s guidance supports this split, noting that coded workflows provide predictability for well-defined tasks, while agents introduce higher costs and compounding errors unless tested in sandboxes with guardrails.

The Router and Incident Handling

A central router function directs each task based on its assigned lane and approval requirements. If a task is marked as Human or requires approval, it is queued for a person. Workflow tasks execute predefined steps, while Agent tasks run with a restricted set of allowed tools. Crucially, three elements are hardcoded outside the prompt: the specific tools a task may access, the approval requirements, and the logging protocol. To handle failures, the system includes a global pause set. When an incident occurs, a named owner must trigger a pause, which is logged with the operator’s identity, ensuring accountability during outages.

Red Lines and Limitations

Pandya is explicit about the boundaries of this system. Money, contracts, and customer promises are red lines that must always route to a human. Every agent action generates a log entry capturing inputs and outputs, and every task has a named owner with an off switch. This is not a proof of concept for a fully self-running company; rather, it operationalizes the "hybrid reasoning" described by Salesforce and the human-in-the-loop direction setting noted in Microsoft’s 2025 Work Trend Index. The control plane transforms vague concepts of AI oversight into executable code.

Key Takeaways

  • Autonomy is a control-plane problem, not just a model selection problem.
  • Three lanes (Workflow, Agent, Human) enforce predictability and safety.
  • Hardcoded routing logic and logging are critical for incident accountability.
  • High-stakes decisions like pricing and contracts must remain human-controlled.

The Bottom Line

Stop optimizing for model intelligence and start optimizing for control plane rigidity. True autonomy in business workflows is defined by what you refuse to let the AI decide, not what you ask it to guess.