Building agents

Step 1 — pick the model

Start with the smallest model that can do the job. Move up only when you hit a quality wall — small models are dramatically cheaper per request and usually much faster.

• Routing, summarisation, simple Q&A — gpt-4o-mini, claude-3-haiku
• Coding, multi-step reasoning — gpt-4o, claude-3-5-sonnet
• Heavy reasoning, planning — gpt-4-turbo, claude-3-opus

Step 2 — write the system prompt

A good system prompt is short, concrete, and gives the agent a single clear role. Avoid stacking dozens of rules — agents follow short prompts much more reliably.

You are a senior code reviewer. For every diff you receive:

1. Summarise the change in one sentence.
2. Flag any breaking API changes.
3. Suggest specific improvements with file:line references.
4. Output in markdown, no preamble.

Step 3 — pick tools

Foxora ships a small but useful default toolset:

• file — read / write inside the agent’s workspace.
• shell — run shell commands in the workspace.
• web — fetch URLs and search the web.
• http — arbitrary HTTP requests with auth headers.
• memory — read / write a long-term notes file.

For anything custom, register your own tools as JSON schemas — see the Building extensions guide.

Step 4 — choose a trigger

• Chat — user starts every run by typing.
• Schedule — cron-driven, no input needed.
• Webhook — HTTP POST starts the run with a JSON body.
• Event — coming soon: triggered by another agent or extension.

Step 5 — ship and iterate

Save the agent and run it a handful of times with realistic inputs. Watch the tool calls in Usage. The two most common bugs:

• The agent uses the wrong tool for a task — tighten the prompt or remove the tool.
• The agent burns tokens looping — add a stop condition or lower the max steps.