Core Concepts
Agent Profiles
Each agent in OPIDE is a configurable profile with its own model, system prompt, and capability set. Pick the right agent for the task. Switch in one click from the chat panel.
The four built-in profiles#
| Profile | Specialty | Best for |
|---|---|---|
| Coder | coder | Writing and editing code, refactors, bug fixes, generation |
| Researcher | researcher | Reading docs, searching the web, understanding APIs and libraries |
| Security | security | Vulnerability review, dependency auditing, threat modelling |
| General | general | Anything that does not clearly fit the other three |
Switching agents#
The active agent is shown in the chat panel header. Click the agent name to open the switcher and select a different profile. One agent runs at a time. The switch takes effect immediately on the next message.
Configuring a profile#
Open Settings (Cmd+,) and navigate to Agents. Each profile exposes three configurable fields:
Model#
Assign any model from any configured provider to this agent. The Coder agent might run on Claude 3.5 Sonnet for its strong code generation, while the Researcher runs on Gemini 2.0 Flash for speed. Model routing falls back through the 5-level hierarchy if no model is set at the agent level.
System prompt#
Each agent has a default system prompt tuned to its specialty. You can append to it or replace it entirely. Useful for injecting project-specific context that every session should know — coding style, architecture decisions, naming conventions.
Trust mode#
Each agent can have its own Trust Mode (Ask, Auto, or Yolo) independent of the global setting. A common setup: Coder on Auto, Security on Ask so every tool call gets reviewed, General on Yolo for exploratory work. See Trust Modes for full details.
How the agent loop works#
Every agent runs the same core execution cycle:
- ◆Assemble context — budget-aware prompt assembly caps system prompt at 45% of context, history at 35% minimum, reply tokens reserved
- ◆Query memory — Engram retrieval pipeline runs, relevant memories are injected into context
- ◆Send to model — streamed through the provider abstraction layer
- ◆Execute tools — tool calls dispatched through the QuickJS sandbox with trust mode enforcement
- ◆Repeat — until a final response or the max round limit is hit
Built-in safety#
- ◆Tool circuit breaker — after 3 consecutive failures on the same tool, the agent gets a warning nudge. After 5, the tool is blocked for the session.
- ◆Repetition detector — hashes tool call signatures and detects loops. Breaks them automatically.
- ◆Yield signaling — if you send a new message mid-turn, the current run is gracefully interrupted.
- ◆Token budget enforcement — the context window is never exceeded. Budget awareness runs at every assembly stage.