Gary Tan's Agentic Workflow Revolution
Gary Tan details his return to building, leveraging agentic AI to ship code at 400x historical rates. This analysis covers the 'token maxing' strategy, the shift from code to markdown skills, and the operational frameworks for high-velocity software development.
The Agentic Shift in Software Engineering
Gary Tan’s return to active building marks a pivotal moment in software development, demonstrating that agentic AI can multiply individual output by hundreds of times. By leveraging tools like Claude Code and OpenClaw, Tan shipped hundreds of thousands of lines of code in months, a volume previously requiring entire teams and years. This case study reveals a new operational paradigm where the primary constraint is no longer coding speed, but the quality of human direction and context management.
Strategic Framework: Token Maxing
A core insight from Tan’s workflow is the concept of "token maxing." Rather than economizing on API usage, the strategy involves spending significantly more on tokens to allow AI models to ingest vast amounts of context. This approach enables the system to cross-reference multiple sources, identify contradictions, and produce outputs that are more complete and accurate than those generated with limited context. Tan compares this to the cost of living in San Francisco: the expense is high, but the cost of not accessing the full utility of the tool is higher. This shifts the budgeting mindset from cost-cutting to value-maximization, treating token spend as a direct investment in product quality.
Architecture: Markdown Over Code
Tan advocates for a structural shift in how software is defined. Complex, context-dependent logic should be encoded in markdown "skills" rather than deterministic code. Code is brittle and fails to handle special cases or user intent, whereas LLMs operating in latent space can interpret nuance and adapt to generic cases. This "thin harness, fat skills" approach allows for more flexible and robust agentic systems. Additionally, requiring agents to generate ASCII diagrams of data flows before coding forces a deeper architectural understanding, reducing bugs and improving the coherence of the final implementation.
Operational Velocity and Quality Control
To maintain quality at high velocity, Tan integrates automated testing directly into the agentic loop. By using browser automation tools like Playwright, the system can perform end-to-end tests without manual intervention. This eliminates the traditional bottleneck of manual QA, allowing for continuous deployment. However, Tan emphasizes that human agency remains irreplaceable. The human role shifts from writing code to defining the "platonic ideal" of the product, providing taste, and making strategic decisions. The machine executes, but the human directs.
Conclusion
The future of software engineering lies in personal AI infrastructure. By building proprietary workflows, developers retain control over their data and prompts, avoiding the opacity of corporate-hosted AI. This era resembles the early days of the personal computer, where individual builders have the power to create sophisticated systems. The key to success is embracing the new tools, investing in context, and maintaining human oversight to ensure that the output aligns with genuine user value.
Key insights
-
Token expenditure should be treated as a strategic investment in output quality. High-volume context ingestion allows AI to cross-reference sources and produce more accurate, comprehensive results.
Impact: Companies that adopt token maxing will achieve superior product quality and research depth compared to those treating API costs as a line item to minimize.
-
Markdown is emerging as a superior medium for defining complex, context-dependent logic compared to deterministic code. It leverages LLM latent space to handle edge cases and user intent more effectively.
Impact: Shifting from code to markdown skills will increase the flexibility and robustness of agentic systems, reducing maintenance overhead and improving adaptability.
-
Mandating ASCII diagram generation before code execution forces agents to visualize architecture and dependencies. This step significantly reduces logical errors and improves the completeness of the final implementation.
Impact: Implementing this planning step can reduce bug rates and rework, leading to faster development cycles and higher code quality.
-
Automated browser testing integrated into the agentic loop eliminates the manual QA bottleneck. This allows for continuous, high-volume feature deployment without sacrificing quality.
Impact: Teams can ship features faster and more frequently, gaining a competitive advantage in time-to-market while maintaining high reliability standards.
-
Human agency remains critical for product taste and strategic direction. AI accelerates execution but cannot replace the human judgment required to define value and prioritize features.
Impact: Organizations that clearly define the human role in the agentic workflow will achieve better alignment with user needs and avoid producing low-value outputs.
Action items
-
Implement a token maxing strategy by increasing API budget for critical tasks. Allow the AI to ingest extensive context to improve output accuracy and depth.
Impact: This will lead to higher quality outputs and more comprehensive research, reducing the need for manual fact-checking and refinement.
-
Refactor complex, context-dependent logic from code to markdown skills. Use markdown to define instructions that require nuance and adaptability.
Impact: This will make your agentic systems more flexible and robust, reducing the brittleness associated with deterministic code.
-
Require your AI agents to generate ASCII diagrams of data flows and state machines before writing code. Use this as a planning step to ensure architectural coherence.
Impact: This will reduce logical errors and improve the completeness of the implementation, leading to fewer bugs and less rework.
-
Integrate browser automation tools like Playwright into your agentic workflow for automated end-to-end testing. Eliminate manual QA steps where possible.
Impact: This will accelerate your deployment cycle and allow for continuous integration and delivery, improving time-to-market.
-
Define clear human oversight roles for product taste and strategic direction. Ensure that humans are involved in defining the 'platonic ideal' of the product.
Impact: This will ensure that the output aligns with user needs and business goals, avoiding the production of low-value or misaligned features.
Quotes
“I think that's like the defining question. Like, will you have control over your own tools or will your tools have control over you?”
“Using OpenClaw these days is like driving a Ferrari and it's like exhilarating. It's insane. Like you get to do things like it figures things out you would never think a machine could figure out and it does it so quickly.”
“Basically, if there is incremental work that... make something more complete, more awesome, more, you know, in the case of this type of writing, like we want it to be more representative of reality.”