README
¶
Complex OpenAI agent example
This example runs a realistic, read-only incident investigation with an OpenAI Responses model. It demonstrates:
- four custom tools with JSON Schema parameters;
- run-scoped region metadata available inside tools;
- built-in task planning and plan updates;
- parallel tool execution with a concurrency limit;
- precise context compression;
- before/after execution callbacks;
- streamed final-answer output and token accounting.
The operational data is deterministic and embedded in the example. No production service is queried or modified.
Run
export OPENAI_API_KEY="your-api-key"
export OPENAI_MODEL="gpt-5.2" # optional; gpt-5.2 is the default
go run ./example/complex_agent
Use a custom question or timeout:
go run ./example/complex_agent \
-question "Investigate the EU checkout errors and recommend a safe mitigation." \
-timeout 2m
OPENAI_BASE_URL can optionally point to an OpenAI-compatible Responses API endpoint.
The agent is intentionally given only read-only tools. Its final response should distinguish observed evidence from hypotheses and recommend, rather than claim to execute, mitigations.
Documentation
¶
Overview ¶
Command complex_agent demonstrates a multi-step OpenAI agent with planning, parallel tool execution, run-scoped metadata, callbacks, context compression, final-answer streaming, and token accounting.
The tools use deterministic in-memory incident data so the example only needs an OpenAI API key and never touches a production system.
Source Files