Build More Effective AI AGENTS With "Code As Action"

“Code as action” lets AI agents do multi-step tool work in one shot by generating and running code (including loops), instead of chaining dozens of separate JSON tool calls. The practical payoff is fewer requests, lower latency, and less tool-call overhead—especially when the same pattern repeats across many inputs.

The walkthrough starts with a cost-optimization example for buying a smartphone model (“code act one”). In a traditional tool-calling approach, the agent runs tools sequentially: it looks up phone prices, converts currencies, estimates taxes, and adds shipping—repeating that whole chain for each candidate country. Even for a small set, the process balloons into many steps and multiple tool calls per country, creating a lot of intermediate requests.

The alternative uses a “code agent” approach where the model writes code that performs the same logic in a loop. Countries are placed into a list, and a for-loop iterates through them to compute the final price for each option. Instead of generating many separate JSON blobs (and running many sequential tool calls), the agent executes the logic as one consolidated code action. The presenter highlights that code actions are more concise than JSON-based action sequences and can collapse what would be “parallel streams” of tool calls into a single step.

Beyond the smartphone demo, the testing shifts to a more analytical task: estimating future API pricing for 2025. A custom setup uses a web tool (via a small agents library) to fetch information like the “DeepSeek R1” API price, “Google Gemini 2.0” API price, and the degree of cost reduction over the last two years. Then the code agent performs calculations—using historical cost reduction (about a 49–50% figure in the run) and applying it to current average prices—to produce projected per-million-token costs for the next year. When errors occur, the agent can retry up to a configured maximum number of steps, feeding error messages back into subsequent code generation.

The final experiment demonstrates extending the agent with a custom tool. Using documentation and an “authorized imports” concept, the presenter builds a tool that uses Matplotlib to visualize projected API price changes and save the output as a PNG file. The code agent then calls this new tool to generate a graph of price per million tokens across models, again emphasizing that the whole workflow can be executed in one action rather than many sequential tool-call steps.

Overall, the key takeaway is operational: “code as action” turns repetitive, multi-tool workflows into loop-driven code execution, which can reduce the number of tool calls, speed up results, and make complex agent behaviors easier to express—while still allowing guardrails through restricted imports for custom tools.