This page documents the Oracle AI Advisor, a specialized sub-agent within Amp's tool ecosystem that provides senior engineering guidance powered by OpenAI's o3 reasoning model. For information about Amp's overall system architecture, see 5.3.1. For details about other sub-agent systems, see Task Tool documentation at 5.3.3 and Codebase Search Agent at 5.3.4.
The Oracle serves as an expert consultant that the main Amp agent invokes for complex reasoning tasks requiring deep analysis, architectural planning, or sophisticated debugging. Unlike execution-focused tools, Oracle specializes in advisory functions, providing recommendations and analysis rather than direct code modifications.
Sources: Amp/gpt-5.yaml220-298 Amp/claude-4-sonnet.yaml211-299
Sources: Amp/gpt-5.yaml253-270 Amp/claude-4-sonnet.yaml211-227
| Property | Value |
|---|---|
| Underlying Model | OpenAI o3 reasoning model |
| Access Method | Tool invocation by main agent |
| Frequency Guidance | "Use FREQUENTLY" (Claude prompt) / "Use for planning, review, debugging" (GPT-5 prompt) |
| User Visibility | Agent mentions Oracle usage: "I'm going to ask the oracle for advice" |
| Response Type | Advisory (recommendations, analysis, guidance) |
Sources: Amp/gpt-5.yaml253-270 Amp/claude-4-sonnet.yaml211-227
Sources: Amp/gpt-5.yaml253-270 Amp/claude-4-sonnet.yaml228-298
The main agent receives explicit guidance on Oracle usage in both Claude and GPT-5 system prompts:
Claude 4.0 Sonnet Prompt:
Use this tool FREQUENTLY. Use it when making plans. Use it to review
your own work. Use it to understand the behavior of existing code. Use
it to debug code that does not work.
Mention to the user why you invoke the oracle. Use language such as
"I'm going to ask the oracle for advice" or "I need to consult with
the oracle."
GPT-5 Prompt:
Senior engineering advisor with o3 reasoning model for reviews,
architecture, deep debugging, and planning.
Use for: Code reviews, architecture decisions, performance
analysis, complex debugging, planning Task Tool runs
Don't use for: Simple file searches, bulk code execution
Prompt it with a precise problem description and attach necessary
files or code. Ask for concrete outcomes and request trade-off
analysis. Use the reasoning power it has.
Sources: Amp/claude-4-sonnet.yaml211-227 Amp/gpt-5.yaml253-270
Sources: Amp/claude-4-sonnet.yaml211-227
| Use Case Category | Specific Examples | Source Prompt |
|---|---|---|
| Architecture | - Review authentication systems - Plan real-time collaboration features - Design JWT implementation | Claude 4.0 Sonnet |
| Debugging | - Race condition analysis - Test failure investigation - Understanding code behavior | Claude 4.0 Sonnet |
| Performance | - Slow query optimization - Performance analysis | Claude 4.0 Sonnet |
| Planning | - Planning Task Tool runs - Strategic feature planning | GPT-5 |
| Review | - Code review - Self-review of agent's work | Claude 4.0 Sonnet |
Sources: Amp/claude-4-sonnet.yaml228-298 Amp/gpt-5.yaml253-270
Sources: Amp/claude-4-sonnet.yaml228-250
Sources: Amp/claude-4-sonnet.yaml239-250
Sources: Amp/gpt-5.yaml220-298
| Feature | Oracle | Task Tool | Codebase Search Agent |
|---|---|---|---|
| Primary Model | OpenAI o3 (reasoning) | Same as main agent | Same as main agent |
| Primary Role | Advisor/consultant | Executor | Explorer |
| Can Modify Files | No (recommendations only) | Yes (direct execution) | No (search only) |
| Reasoning Depth | Deep (o3 specialized) | Standard | Standard |
| Use Frequency | FREQUENTLY | For heavy work | For discovery |
| Best For | Complex decisions | Known implementation | Unknown code location |
| Avoid For | Simple searches | Exploratory work | Code changes |
Sources: Amp/gpt-5.yaml220-298 Amp/claude-4-sonnet.yaml211-299
Recommended workflow pattern: "Oracle (plan) → Codebase Search (validate scope) → Task Tool (execute)"
Sources: Amp/gpt-5.yaml289-298
The system prompts provide specific guidance on how to prompt Oracle effectively:
GPT-5 Prompt Structure:
Claude Prompt Structure:
Sources: Amp/gpt-5.yaml253-270 Amp/claude-4-sonnet.yaml211-227
According to system prompts, do not use Oracle for:
glob or Grep instead)Task tool instead)Grep instead)Sources: Amp/gpt-5.yaml260-286
The system supports parallel Oracle invocations for distinct concerns:
Parallelization rule: "Oracle: distinct concerns (architecture review, perf analysis, race investigation) in parallel."
Sources: Amp/gpt-5.yaml82-114
Oracle configuration is embedded in the system prompt, accessible via:
Alt (Windows) or Option (macOS) and click workspace buttonFor GPT-5-specific Oracle configuration:
Sources: Amp/README.md1-21
Oracle documentation appears in the system prompt text at:
The tool is referenced but not defined as a separate tool schema, indicating it's a backend service invoked through the main agent's reasoning process.
Sources: Amp/claude-4-sonnet.yaml1-1454 Amp/gpt-5.yaml1-2287
The Claude system prompt emphasizes: "Use this tool FREQUENTLY" - indicating Oracle is not a last resort but a primary reasoning amplifier for the main agent.
The agent must announce Oracle usage: "I'm going to ask the oracle for advice" or "I need to consult with the oracle" - maintaining user awareness of the enhanced reasoning step.
Oracle prompts should "Ask for concrete outcomes" rather than abstract advice, ensuring actionable recommendations.
Oracle's distinction comes from using OpenAI's o3 reasoning model specifically, which provides enhanced analytical capabilities compared to the main agent's model (Claude 4.0 Sonnet or GPT-5).
Sources: Amp/claude-4-sonnet.yaml211-227 Amp/gpt-5.yaml253-270
Refresh this wiki