Orchestrating AI Agents for Rapid Prototyping
This past weekend, a developer tackled the challenge of building a functional AI tool for Google's All Things Agentic Hackathon. The project, named Setback, aims to streamline the process for residents objecting to local development applications. Instead of manually crafting submissions that often include legally irrelevant points, Setback converts plain-language objections into formal council-ready documents. Crucially, it also identifies and rejects arguments that fall outside the scope of planning law, ensuring submissions are both effective and compliant.
The core innovation lies not just in the tool's functionality but in its development process. The bulk of the coding was handled not by the developer directly, but by AI coding agents under their orchestration. This approach allowed for rapid prototyping, demonstrating the power of AI agents in accelerating software development, particularly in time-constrained environments like hackathons.
The problem Setback addresses is a common one in local governance. In New South Wales, planning law specifies a narrow list of considerations for councils reviewing development applications. These typically include factors like overshadowing, privacy, building height and bulk, and heritage impact. Personal preferences, such as disliking the applicant or concerns about falling property values, are not legally admissible grounds for objection, yet residents frequently include them in their submissions. This dilutes the impact of valid concerns and complicates the review process.
The Technical Architecture of Setback
Setback's architecture is built around a central orchestrator that manages several specialized AI agents. The process begins when a resident provides their objection in natural language. This raw input is then processed by an agent tasked with understanding and parsing the user's intent and concerns. This agent acts as a preliminary filter, identifying the core issues the resident wishes to raise.
Following this initial parsing, the information is fed to a legal research agent. This agent's role is to cross-reference the resident's stated concerns against the legally permissible grounds for objection under NSW planning law. It queries knowledge bases and legal precedents to determine which parts of the objection are relevant and which are not.
The next stage involves a content generation agent. This agent takes the filtered, legally relevant points and crafts them into a formal submission document. It adheres to the required structure and language expected by local councils. Simultaneously, a separate agent handles the formulation of polite but firm refusals for any points deemed legally irrelevant. This ensures that the resident is informed about why certain objections were not included in the formal submission, maintaining transparency.
Finally, an overarching orchestrator agent coordinates these specialized agents, managing the flow of information between them, handling errors, and ensuring the final output is coherent and complete. This agent is akin to a project manager, assigning tasks to its sub-agents and integrating their results. The developer's role was primarily to define the overall workflow, set the parameters for each agent, and troubleshoot any integration issues.
Challenges and Surprises in Agent Orchestration
One of the most significant challenges was ensuring seamless communication and data transfer between the different AI agents. Each agent was designed to perform a specific function, but integrating their outputs into a cohesive whole required careful prompt engineering and API management. For instance, ensuring that the output format of the legal research agent perfectly matched the input requirements of the content generation agent was critical.
The surprising detail here was not the speed at which the prototype was built, but the level of nuance the AI agents could handle. Initially, the expectation was that the agents might struggle with the subjective nature of resident objections and the fine distinctions in planning law. However, with well-defined prompts and access to relevant data, the agents demonstrated a remarkable ability to differentiate between legally valid concerns and personal opinions, and to articulate this distinction clearly.
The developer found that the AI agents acted less like independent workers and more like specialized tools that needed precise instructions. The art of orchestrating them involved understanding each agent's strengths and limitations, and designing the workflow to leverage those strengths while mitigating weaknesses. This is a fundamentally different paradigm from traditional software development, where developers write every line of code themselves.
Broader Implications for AI Development and Governance
The Setback project highlights a powerful paradigm shift in software development. By orchestrating AI agents, developers can move from writing code line-by-line to designing workflows and managing AI collaborators. This dramatically reduces development time, making complex applications feasible within extremely short periods, such as a weekend hackathon.
For governance, tools like Setback could significantly improve the quality and relevance of public submissions. By ensuring that objections are legally sound, councils can process them more efficiently, leading to better-informed development decisions. This also empowers residents by providing them with a tool that helps articulate their concerns effectively within the established legal framework.
However, this approach also raises questions about accountability. When AI agents are responsible for generating formal legal documents, where does the ultimate responsibility lie? The developer acted as the orchestrator, but the AI performed the actual writing. This dynamic will become increasingly important as AI agents are deployed in more critical applications, from legal and medical fields to financial services.
The success of Setback in a hackathon setting demonstrates the viability of agent-based development for rapid prototyping and problem-solving. It suggests a future where complex software solutions can be assembled from specialized AI components, managed by a human orchestrator, opening new avenues for innovation and efficiency.