Agents¶

Morphix includes five built-in agents, each with a distinct role, personality, and tool set. Agents are the workers that execute subtasks in workflows — they reason about your request, use tools to interact with your project, and produce structured output.

Agent Registry¶

Agents are loaded from YAML templates in templates/agents/ (copied to workspaces/<name>/agents/ on first workspace switch). Each agent has a profile with:

System prompt: Instructions, personality, constraints
Tools: Allowed tool list
Keywords: Used for routing (agent selection based on task description)
Model role: Which LLM role to use (agent, reasoning, fast)
Temperature: Response randomness (0.0 = deterministic, 1.0 = creative)
Priority: Routing priority (higher = preferred for ambiguous tasks)

Developer¶

Property	Value
Type	`development` (agent)
Model role	`agent`
Temperature	0.2
Priority	70
Tools	`file_manager`, `git_manager`, `bash_manager`, `lsp_manager`, `code_exec`, `test_runner`, `diff_editor`

The Developer agent writes, modifies, and tests code. It has the most extensive tool set of any agent — it can read/write files, run shell commands, execute tests, apply diffs, and manage Git.

Best Use Cases¶

Writing new code (functions, classes, modules, APIs)
Fixing bugs and refactoring
Running tests and debugging failures
Installing dependencies via bash
Creating commits

Constraints¶

Uses relative paths only — never absolute paths or code_projects/ prefix
bash_manager requires the command parameter (without it, the tool fast-fails)
python3 -c ... is blocked by security; use test_runner instead
code_exec sandbox blocks os, sys, subprocess, io, socket, requests, pathlib

Example Prompt¶

Add pagination to the /api/users endpoint in the Flask app.
The endpoint should accept ?page=1&per_page=20 query parameters.
Return a JSON response with "data", "page", "per_page", and "total" fields.

Analista¶

Property	Value
Type	`analysis` (reasoning)
Model role	`reasoning`
Temperature	0.2
Priority	55
Tools	`file_manager`, `lsp_manager`, `code_search`, `web_search`

The Analista agent analyzes and reviews — it reads code, evaluates architecture, identifies risks, and provides recommendations. It is read-only: it can read files, search code, and navigate with LSP, but it cannot write files or commit.

Best Use Cases¶

Code review and security analysis
Architecture evaluation
Identifying bugs, performance issues, anti-patterns
Explaining what code does
Recommending improvements

Constraints¶

No file_manager.write, git_manager.commit, or diff_editor
Read-only analysis — hands off implementation to the developer agent
Output is structured analysis, not code

Example Prompt¶

Review the authentication module in src/auth/ for security issues.
Look for: SQL injection, hardcoded secrets, missing input validation,
insecure password handling, and CSRF vulnerabilities.

Moderador¶

Property	Value
Type	Agent (no specific type)
Model role	`reasoning`
Temperature	0.4
Priority	1 (lowest — never auto-selected)
Tools	none

The Moderador agent is the neutral arbiter for the Collaborative workflow. It listens to all panelists' arguments, identifies common ground, guides toward consensus, and delivers the final verdict. It has no tools — its value is in synthesis and balanced judgment.

Best Use Cases¶

Synthesizing a debate into a final consensus
Weighing conflicting opinions fairly
Delivering structured, balanced conclusions

Constraints¶

No tools — pure reasoning agent
Lowest priority — never selected by the AgentRouter for code tasks

Example Prompt¶

(This agent is automatically invoked by the Collaborative workflow.
You don't call it directly.)

Conversacional¶

Property	Value
Type	Agent (no specific type)
Model role	`agent`
Temperature	0.4
Priority	10
Tools	none

The Conversacional agent is the friendly conversationalist. It handles small talk, greetings, profile questions, and general help. It has no tools — it's a pure chat agent designed for natural conversation.

Best Use Cases¶

Greetings and casual conversation
"What can you do?" / "Help me get started"
Profile and preference questions
Quick factual explanations

Example Prompt¶

Hi! What can you help me with today?

Architect¶

Property	Value
Type	`analysis` (reasoning)
Model role	`reasoning`
Temperature	0.2
Priority	58
Tools	`file_manager`, `lsp_manager`, `code_search`, `web_search`

The Architect agent designs system structure before implementation. It maps existing code, evaluates patterns and libraries, and produces clear specifications with an ordered implementation plan. Like Analista, it is read-only — it designs, but does not write production code.

Best Use Cases¶

Designing system architecture for new features
Planning module boundaries and interfaces
Choosing between frameworks/libraries with trade-off analysis
Creating implementation plans for the developer agent

Constraints¶

No file_manager.write, git_manager, or diff_editor
Delivers specifications and plans — hands off to developer for implementation

Example Prompt¶

Design the architecture for a notification system that supports:
- Email, SMS, and push notifications
- User preferences per channel
- Async delivery with retry logic
- Rate limiting per user

Provide component responsibilities, data flow, and an implementation plan.

Agent Routing¶

When a task is decomposed, the AgentRouter selects the best agent for each subtask based on:

Keyword matching — Each agent has a keywords list in its profile. The router matches these against the subtask description.
Priority — Higher-priority agents are preferred when multiple match.

Agent	Priority	Key keywords
Developer	70	code, implement, create, build, deploy, refactor, fix, test, debug, api, endpoint
Architect	58	architecture, design, structure, blueprint, plan, components, interfaces, patterns
Analista	55	analyze, review, evaluate, architecture, risks, security, performance, diagnose
Conversacional	10	hello, help, thanks, small talk, what can you do
Moderador	1	consensus, moderate, debate, discuss (only used in Collaborative)

Force an agent

You can override auto-routing by selecting a specific agent in the Maestro agent dropdown. In Orchestrate mode, only agents allowed by the active workflow are shown. In Chat mode, all agents are available.