Agent

Agent — AI Agent Development

Activate When

• User invokes /godmode:agent
• User says "build an AI agent", "create an agent", "add tools to my agent"
• User says "design agent memory", "agent keeps looping", "agent safety"
• When building autonomous or semi-autonomous LLM-powered systems
• When /godmode:prompt identifies a need for agentic capabilities (tool use, multi-step reasoning)
• When /godmode:rag needs to be wrapped in an agent loop
• When the orchestrator detects agent frameworks (LangChain, LlamaIndex, CrewAI, AutoGen, custom agent loops) in code

Workflow

Step 1: Agent Discovery & Requirements

Understand what the agent must accomplish:

AGENT DISCOVERY:
Purpose: <what the agent must autonomously accomplish>
Type:
  - Single-agent: One agent with tools (most common)
  - Multi-agent: Multiple specialized agents coordinating
  - Human-in-the-loop: Agent proposes, human approves critical actions

User interaction:
  - Conversational: User chats with agent in real-time
  - Autonomous: Agent runs a task to completion without user input
  - Supervised: Agent asks for confirmation at decision points

Environment:
  - Tools available: <list of APIs, databases, code execution, file systems>
  - External systems: <services the agent will interact with>

If the user hasn't specified, ask: "What should this agent do autonomously? What tools does it need?"

Step 2: Architecture Pattern Selection

Select the agent architecture pattern:

AGENT ARCHITECTURE SELECTION:

Patterns:
| Pattern | Best for |
|--|--|
| ReAct | General-purpose tool use, step-by-step reasoning |
| (Reason + Act) | with tool calls. Simple, effective, well-understood. |
| Plan-and-Execute | Complex tasks needing upfront planning. Planner |
|  | creates step list, executor follows it. Good for |
|  | multi-step tasks with clear decomposition. |
| Reflexion | Tasks requiring self-correction. Agent attempts, |
|  | evaluates own output, and retries with feedback. |

Step 3: Agent Loop Design

Design the core agent execution loop:

AGENT LOOP DESIGN:

Pattern: <selected pattern>
Model: <LLM for agent reasoning — e.g., Claude 3.5 Sonnet, GPT-4>

ReAct loop:
  while not done and steps < max_steps:
  1. THINK: Reason about current state and what to do next
  2. ACT: Select and call a tool with parameters
  3. OBSERVE: Process tool result
  4. EVALUATE: Is the task complete? Should I continue?
  Termination conditions:
  - Task completed successfully -> return result

Step 4: Tool Design & Integration

Design the tools the agent can use:

TOOL INVENTORY:
| Tool | Type | Risk Level | Description |
|--|--|--|--|
| <tool_name> | Read-only | LOW | <what it does> |
| <tool_name> | Write | MEDIUM | <what it does> |
| <tool_name> | External API | MEDIUM | <what it does> |
| <tool_name> | Code exec | HIGH | <what it does> |
| <tool_name> | Destructive | CRITICAL | <what it does> |

TOOL DESIGN PRINCIPLES:
1. Single responsibility: each tool does one thing well
2. Clear naming: tool name describes the action (search_docs, create_ticket)
3. Typed parameters: every parameter has a type, description, and constraints

Step 5: Memory System Design

Design how the agent remembers and learns:

MEMORY SYSTEM DESIGN:

Memory types:
| Type | Implementation |
|--|--|
| Working memory | Current conversation context window. Limited by |
| (short-term) | model context length. Contains current task state, |
|  | recent tool results, and immediate reasoning. |
| Conversation memory | Full conversation history, summarized on overflow. |
|--|--|
| (session) | Stored in session store (Redis, database). |
|  | Summarize older turns to fit context window. |
| Episodic memory | Past task executions and outcomes. "Last time I |

Step 6: Guardrails & Safety

Design safety boundaries for the agent:

AGENT GUARDRAILS:

Layer 1 — Input guardrails:
| Check | Action |
|--|--|
| Prompt injection detection | Reject input, log attempt |
| PII in input | Redact before processing |
| Off-topic request | Redirect to designated channel |
| Malicious intent detection | Refuse and log |
| Input length limit | Truncate with warning |

Layer 2 — Execution guardrails:

Step 7: Agent Evaluation & Testing

Design a test suite for the agent:

AGENT TEST SUITE:

Test categories:
| Category | Tests | Description |
|--|--|--|
| Task completion | <N> | Agent successfully completes defined tasks |
| Tool selection | <N> | Agent picks correct tool for each step |
| Multi-step reasoning | <N> | Agent chains tools correctly for complex |
|  |  | tasks |
| Error recovery | <N> | Agent handles tool failures gracefully |
| Safety compliance | <N> | Agent refuses unsafe actions |
| Guardrail adherence | <N> | Agent stays within defined limits |
| Edge cases | <N> | Ambiguous inputs, missing data, conflicts |
| Adversarial | <N> | Injection attacks, manipulation attempts |

Step 8: Agent Artifacts & Commit

Generate the deliverables:

1. Agent config: config/agents/<agent>-config.yaml
2. Agent implementation: src/agents/<agent>/agent.py
3. Tool definitions: src/agents/<agent>/tools/
4. Memory module: src/agents/<agent>/memory.py
5. Guardrails: src/agents/<agent>/guardrails.py
6. Test suite: tests/agents/<agent>/
7. Architecture doc: docs/agents/<agent>-architecture.md

AGENT DEVELOPMENT COMPLETE:

Architecture:
- Pattern: <pattern name>
- Model: <LLM for reasoning>
- Tools: <N tools> (read: <N>, write: <N>, code exec: <N>)
- Memory: <memory types implemented>
- Guardrails: <N guardrail layers>

Evaluation:
- Task completion rate: <val>
- Tool selection accuracy: <val>
- Safety violation rate: <val> (require 0%)
- Avg steps per task: <N>
- Avg latency per task: <seconds>

Commit: "agent: <agent name> — <pattern>, <N> tools, completion=<val>, safety=100%"

Key Behaviors

1. Guardrails before capabilities. Safety constraints first.
2. Tools are the agent's hands. Tool design > LLM choice.
3. Loops need escape hatches. Max steps, cost, time limits.
4. Test trajectories, not only outcomes. Evaluate full traces.
5. Memory is context engineering. Retrieve selectively.
6. Human-in-the-loop for irreversible actions. No exceptions.
7. Observability is mandatory. Log every agent step.

# Run agent evaluation suite
pytest tests/agents/ -v --timeout=120
python -m agents.evaluate --test-inputs 3 --safety-check

IF completion rate < 80%: review tool definitions and prompts. WHEN safety violation detected: block deployment, fix immediately.

Flags & Options

Flag	Description
(none)	Full agent development workflow
--pattern <name>	Force architecture: react, plan-execute, reflexion, multi-agent, state-machine, router
--tools	Design and inventory agent tools

Explicit Loop Protocol

When building or debugging agent loops, use this tracking protocol:

AGENT BUILD/DEBUG LOOP:
current_iteration = 0
max_iterations = 20
issues_remaining = total_issues

WHILE issues_remaining > 0 AND current_iteration < max_iterations:
    current_iteration += 1

    1. IDENTIFY next issue (tool gap, guardrail weakness, test failure)
    2. IMPLEMENT fix (code change, config update, prompt edit)
    3. git commit with message: "agent: fix <issue> (iter {current_iteration})"
    4. RUN evaluation suite against the fix
    5. RECORD result:
       - Pass/fail for each test category
       - Regression check: did the fix break anything?

HARD RULES

MECHANICAL CONSTRAINTS — NON-NEGOTIABLE:
1. NEVER deploy an agent without guardrails defined first — safety before capabilities.
2. NEVER allow irreversible tool actions without explicit user confirmation gate.
3. EVERY agent loop MUST have a max_steps termination — no unbounded loops.
4. EVERY agent MUST have a cost budget (max tokens per task) — no runaway spending.
5. git commit BEFORE running evaluation — if eval reveals regression, revert.
6. Safety violation rate MUST equal 0% — any safety failure is a blocking issue.
7. Log every agent step in structured format:
   STEP\tACTION\tTOOL\tRESULT\tTOKENS\tLATENCY
8. Test trajectories, not only final outputs — correct answer via unsafe path is a failure.
9. NEVER give agents tools they do not need — fewer tools = better tool selection.
10. Observability is mandatory — if you cannot trace every step, do not deploy.

Auto-Detection

AUTO-DETECT agent context:
  1. LLM provider: grep -r "openai\|anthropic\|google.generativeai\|ollama\|together" package.json pyproject.toml
    requirements.txt 2>/dev/null
  2. Agent framework: grep -r "langchain\|langgraph\|autogen\|crewai\|magentic\|pydantic-ai" package.json
    pyproject.toml 2>/dev/null
  3. Tool definitions: grep -rl "tool_call\|function_call\|@tool\|BaseTool\|StructuredTool" src/ --include="*.ts"
    --include="*.py" 2>/dev/null | head -5
  4. Vector store: grep -r "pinecone\|weaviate\|chromadb\|pgvector\|qdrant\|milvus" package.json pyproject.toml
    2>/dev/null
  5. Existing agent code: grep -rl "agent\|AgentExecutor\|ReActAgent\|create_agent" src/ --include="*.ts"
    --include="*.py" 2>/dev/null | head -5

Success Criteria

Verify all of these before marking the task complete:

1. Agent completes target task on >= 3 test inputs.
2. Max steps termination works (limit <= 20 steps).
3. Cost budget enforced (token limit per task <= 100K tokens).
4. Guardrails block unsafe actions (>= 1 adversarial test).
5. Every step logged: step, action, tool, result, tokens, latency.
6. Irreversible actions require confirmation gate.
7. Tool errors handled gracefully (max 2 retries).
8. Evaluation suite measures success rate, cost, safety.

<!-- tier-3 -->

Error Recovery

Failure	Action
Agent loops without progress	Add loop detection: if same action repeated 3x, force different action or stop.
Token budget exceeded	Set hard limit per task. When 80% consumed, switch to shorter prompts.
Tool returns unexpected format	Validate output. Retry max 2x, then report failure.

Keep/Discard Discipline

After EACH agent change (prompt edit, tool addition, guardrail update):
  1. MEASURE: Run evaluation suite — task completion rate, safety violations, avg steps.
  2. COMPARE: Did the change improve the target metric without introducing regressions?
  3. DECIDE:
     - KEEP if: completion rate maintained or improved AND safety violations = 0 AND no new failure modes
     - DISCARD if: safety violation detected OR completion rate dropped OR new failure mode introduced
  4. COMMIT kept changes. Revert discarded changes before the next iteration.

Never keep a change that introduces any safety violation, regardless of completion rate improvement.

Stop Conditions

STOP when ANY of these are true:
  - Agent completes target tasks end-to-end with correct output on 3+ test inputs
  - Safety violation rate = 0% across all test cases including adversarial inputs
  - All guardrails (max steps, cost budget, confirmation gates) verified working