Field & Signal

Inspiration

Small businesses make important decisions about customers, pricing, products and expansion, but proper market research can be expensive and slow. AI has made web research easier, yet most tools still only search what is already known. Field & Signal began with a simple question: What if a small business could engage an entire AI-native market research agency as easily as describing its decision?

What it does

Field & Signal turns a business question into a complete research engagement led by six clearly disclosed AI specialists. The team: 1) Clarifies the business decision. 2) Produces an approval-ready research plan. 3) Finds and records relevant public evidence. 4) Creates a focused survey and interview guide. 5) Hosts genuine surveys and consent-based adaptive interviews. 6) Integrates the evidence into a decision-ready brief.

Each specialist has a distinct role. John plans the research, Maya investigates public evidence, Aisha designs the methodology, Daniel conducts interviews, Sofia integrates the findings and Marcus prepares the recommendation.

The application keeps client approval, participant consent, evidence links and research limitations visible throughout the process.

How we built it

Field & Signal is built with Next.js, TypeScript, Supabase, the OpenAI Responses API, GPT-5.6 and Vercel.

I used ChatGPT during ideation to develop the product concept, agent responsibilities and user journey. These discussions helped me recognise that primary research—not simply having multiple agents—should be the product’s main differentiator. I made the final product, methodology and design decisions. I then collaborated extensively with Codex to turn those decisions into a working application. Codex helped implement the Next.js and Supabase architecture, connect GPT-5.6 through the Responses API, build the interface, generate visual assets and improve responsive behaviour. During testing, Codex also diagnosed malformed model outputs, irrelevant research results, database issues and broken workflow transitions. GPT-5.6 powers the six specialist agents. It develops research plans, designs research instruments, generates adaptive interview follow-ups, integrates evidence and produces the final brief.

The process was iterative: I tested each deployment on Vercel, identified problems from a client’s perspective and worked with Codex to implement, verify and deploy each improvement.

Challenges we ran into

One major challenge was avoiding “multi-agent theatre.” The specialists needed to perform real, distinct work rather than simply respond with different personalities. I solved this by creating persistent research artefacts and clear handoffs between planning, evidence gathering, fieldwork, analysis and reporting. Another challenge was making model output reliable. Research plans sometimes returned malformed JSON or unexpected field types. Structured schemas, validation, normalisation and bounded question counts made the workflow more dependable.

Live research and analysis can also take time, so I created progress interfaces that explain which specialist is working and what is happening.

Most importantly, I learned that useful autonomy does not mean removing humans from every decision. Field & Signal allows the agents to operate the research workflow while keeping external actions approval-controlled and research participants informed.

Accomplishments that we're proud of

We built a functioning end-to-end research workflow rather than a collection of static agent demonstrations. Field & Signal can turn a business question into an approved research plan, conduct live secondary research, publish a real survey, store genuine responses, run consent-based adaptive interviews and produce an integrated, traceable brief. We are especially proud that the six GPT-5.6 specialists perform distinct jobs and pass persisted research artefacts between them. We also created clear progress indicators, approval controls and evidence links so users can understand what the agents are doing and where the recommendation came from.

What we learned

We learned that useful autonomy is not about removing humans from every decision. It is about allowing agents to manage the operational work while keeping important actions, such as approving a research plan or beginning fieldwork, under human control.

We also learned that trustworthy AI research requires more than a strong model. Structured validation, relevant sources, clear methodological limitations and thoughtful loading and error states are equally important. Iterative testing with Codex helped us turn model and interface failures into a more dependable workflow.

What's next for Field & Signal

Next, we would expand how Field & Signal recruits and reaches respondents through approved panel providers, client-supplied contact lists and targeted recruitment campaigns.

We would also add multilingual voice interviews, richer survey analysis, respondent-quality checks, document uploads and continuous monitoring of the assumptions behind a recommendation.

The longer-term goal is to make rigorous primary and secondary research accessible to smaller organisations without requiring them to assemble or engage a traditional specialist team.

Built With

• chatgpt
• codex
• gpt-5.6
• next.js
• openai
• supabase
• the-openai-responses-api
• typescript
• vercel