How to Build Any Specialized AI Agent — CEO, HR, Sales, Support — Using the AACFlow Agent Block

Tools — attach any of the 300+ tools and integrations available on the platform. Firecrawl for web scraping, Perplexity for research, Slack for notifications, Gmail for email, and hundreds more. The tool-input sub-block lets you select which tools the agent has access to, with per-tool usage control (auto/required/none).

Skills — attach reusable skill configurations that extend agent capabilities without redefining them every time.

Memory — four modes:

• None: stateless execution, each run is independent.
• Conversation: persistent memory keyed by conversationId. Every interaction with the same ID accumulates context. A customer support agent remembering the full history of a ticket.
• Sliding window (messages): keeps the last N messages, defined by slidingWindowSize. Cost-controlled, deterministic token usage.
• Sliding window (tokens): keeps messages up to a token budget, defined by slidingWindowTokens. More precise cost control for token-based billing.

Temperature — a slider from 0 to 1 (or 0 to 2 for models that support extended ranges). Default is 0.3, which balances creativity with reliability.

Max Output Tokens — cap the response length. Prevents runaway costs on long generations.

Response Format — a JSON Schema that constrains the agent's output. When present, the agent returns structured data instead of free text. Fields are accessible downstream as <agent.field>. This is how you build deterministic pipelines: the agent doesn't just "write something" — it returns a predictable data structure that the next block consumes.

Previous Interaction ID — reference a prior agent execution to continue a conversation thread, enabling multi-turn workflows.

The block outputs: content, model, tokens (prompt/completion/total), toolCalls (list and count), providerTiming, cost, and interactionId.

Building a CEO Agent

A CEO agent needs strategic reasoning, multi-source research synthesis, and the ability to produce decision-quality memos. Here's the exact configuration:

System Prompt (generated by wandConfig AI):

Model: Claude Opus 4.5 (for deep reasoning) or GPT-4.1 (for structured output) Tools: Perplexity (research), Firecrawl (web scraping), Browser Use (interactive research), Slack (notification) Memory: Conversation mode with conversationId: ceo-daily-briefing — builds institutional knowledge over time Response Format: JSON Schema with fields for executiveSummary, strategicAnalysis, riskAssessment, recommendations, assumptions, decisionFramework Temperature: 0.2 (low — strategic analysis needs consistency, not creativity)

Real workflow: Every Monday at 8 AM, a trigger fires → CEO Agent gathers market data from 12 sources via Firecrawl + Perplexity → analyzes competitive moves → generates a structured briefing memo → posts to Slack. The CEO starts the week with an AI-generated strategic overview that took 45 seconds to produce, not 4 hours.

Building an HR Agent

An HR agent screens candidates, generates interview questions tailored to the role, and drafts offer letters. It needs structured thinking and compliance awareness.

System Prompt:

Model: GPT-4o (balanced speed and quality) Tools: Gmail (send offer letters), Slack (notify hiring manager), Google Sheets (update candidate tracker) Memory: Conversation mode with conversationId: candidate-{candidateId} — tracks the entire hiring pipeline for each candidate Response Format: Structured JSON for ATS integration

Building a Sales Agent

A sales agent qualifies leads, researches prospects, crafts personalized outreach, and manages follow-up sequences.

System Prompt:

Model: Claude Sonnet 4 (excellent at persuasive writing with factual precision) Tools: Firecrawl (company research), Hunter.io (email finding), HubSpot/Salesforce (CRM integration), Slack (deal alerts) Memory: Conversation mode with conversationId: deal-{dealId} — tracks the full negotiation history Temperature: 0.5 (slightly higher — outreach needs natural language variation)

The wandConfig advantage

Building these system prompts from scratch is hard. The Agent block's wandConfig AI eliminates that friction. You describe the role in plain language — "I need an agent that reviews legal contracts for compliance with GDPR" — and it generates a complete system prompt with methodology, output structure, quality standards, and edge-case handling. It maintains editing history, so you can iterate: "add a section for data processing agreement review" and it updates the existing configuration intelligently.

The wandConfig prompt template is itself a carefully engineered meta-prompt that teaches the AI how to design agent personas. It includes three reference examples — Research Specialist, Senior Code Reviewer, and Writing Editor — that demonstrate the expected depth and structure.

The foundation, not the ceiling

The Agent block is the atomic unit of AACFlow's agent system. Every specialized agent — whether it's a CEO analyst, an HR screener, or a sales SDR — starts with the same block. What makes them different is how you configure it: the system prompt gives it identity, the tools give it capabilities, the memory gives it continuity, and the response format gives it determinism.

And when one agent isn't enough — when you need a team — that's where the Mothership orchestrator comes in. But that's for the next post.

— Alexandr Chibilyaev