A developer I follow described his AI setup like this: "I have three work agents: the administrative assistant, the researcher, and the chief of staff. The chief of staff simulates my boss's communication style, and I use it to practice and polish presentations. For personal agents, I have a life coach, a content and expression assistant, and a finance analyst."
That's six distinct roles running on the same model, differentiated entirely by system prompts and context files. No custom software. No platform subscription beyond the API. Just instructions that shape how the model responds.
This is where the real value of tools like Claude Code, Codex, and Managed Agents lives. Not in single-shot coding tasks, but in building a staff of specialized agents that handle recurring work the way a human hire would, except they cost tokens instead of salaries and they're available at 2am.
The five roles and how to configure them
1. The administrative assistant
What it handles: Email drafts, meeting agendas, calendar summaries, document formatting, follow-up reminders, status reports.
How to configure it: Create a system prompt that includes your communication style, your standard meeting format, your email signature, and your recurring commitments. Store this as a Markdown file (see the Markdown memory pattern) so the agent loads it automatically at the start of every session.
In Claude Code, this means a CLAUDE.md that says: "You are my administrative assistant. My communication style is [specific examples]. When drafting emails, match this tone. My standard meeting agenda format is [template]. When I say 'draft a follow-up,' generate an email based on the last conversation context."
In Codex, the equivalent goes in the project instructions file. The model reads it at session start and shapes every response accordingly.
The pattern that works: End-of-day summaries. Ask the agent to compile what happened today (from your notes, git log, calendar) into a one-paragraph summary and three action items for tomorrow. Takes two minutes. Saves twenty minutes of mental overhead the next morning.
2. The researcher
What it handles: Competitive analysis, market research, documentation lookups, citation gathering, fact-checking claims before you publish them.
How to configure it: The researcher agent needs web access. In Claude Code, this means MCP tools or the built-in web search. In Gemini CLI, the grounded search is built in. In Codex, you'd pipe the research task and review the output.
The system prompt should specify your research standards: primary sources preferred over blog posts, specific statistics over vague claims, and always include the source URL so you can verify. If you write blog posts or reports, add your citation format to the prompt.
The pattern that works: Before writing anything data-heavy, ask the researcher agent to verify every statistical claim in your draft. "Check each number in this document. For each one, find the original source and confirm the number is accurate. Flag anything you can't verify." This catches errors before publication.
3. The chief of staff
What it handles: Presentation review, communication coaching, decision preparation, stakeholder simulation.
How to configure it: This is the most interesting role because it requires the agent to simulate someone other than you. Feed it examples of how your manager, your board, or your key stakeholder communicates. Include their priorities, their pet peeves, their decision-making patterns.
"You are simulating [role]. When I present an idea, push back the way [this person] would. Ask the questions they'd ask. Flag the concerns they'd raise. Grade my presentation on clarity, persuasiveness, and whether I addressed their known priorities."
The pattern that works: Practice a presentation or pitch three times against the simulated stakeholder before delivering it live. Each round, the agent pushes back differently based on your responses. By the time you're in the real meeting, you've already heard the hard questions.
4. The finance analyst
What it handles: Budget tracking, expense categorization, revenue projection, pricing analysis, cash flow modeling.
How to configure it: The finance agent needs your actual numbers. Create a Markdown file with your current revenue, expenses by category, pricing tiers, and financial targets. Update it monthly. The agent reads this context and does analysis against it.
"You are my finance analyst. My current monthly revenue is [X]. My expenses by category are [table]. When I ask about pricing changes, model the impact on margin. When I ask about a new expense, tell me what revenue it needs to justify itself."
The pattern that works: Monthly finance review. Paste your bank statement summary or accounting export. Ask the agent to categorize expenses, flag anomalies, calculate month-over-month changes, and project the next quarter based on current trends. What used to take an afternoon with a spreadsheet takes twenty minutes with context-loaded prompts.
5. The content assistant
What it handles: Blog drafts, social posts, newsletter copy, product descriptions, SEO metadata.
How to configure it: This agent needs your voice. Include writing samples, your banned-word list, your formatting preferences, and your audience profile in the context file. The more specific you are about what you don't want ("no em-dashes, no pitch-deck language, no slop words"), the cleaner the output.
"You are my content assistant. My voice samples are [attached]. Never use these words: [list]. My audience is [description]. When drafting a blog post, follow this structure: [template]. Always include a fact-check notes section for data-heavy posts."
The pattern that works: Write the outline yourself. Let the agent draft from the outline. Edit the draft yourself. This division of labor keeps your voice in the work while offloading the mechanical writing. The agent handles the first 80%; you handle the last 20% that makes it yours.
The infrastructure
All five agents run on the same model. The differentiation is entirely in the system prompt and context files. You can run them:
- In Claude Code with different project directories, each containing its own
CLAUDE.mdwith the agent's role and context - In Codex with different instruction files per project
- Via Claude Managed Agents (launched April 2026) where you define the role, tools, and guardrails and Anthropic handles the infrastructure
- Via the Claude Agent SDK if you want to build custom orchestration (the operator pattern, where a controlling agent delegates to specialized sub-agents)
- Via
aichatorllmwith different configuration files per role
The Managed Agents route is the most hands-off. You describe the agent's job, connect the tools it needs (web search, file access, code execution), set the guardrails (what it can and can't do), and Anthropic handles scaling, retries, and context management. For someone running a small business who wants the agent staff without building the plumbing, this is the path.
The key to making any of this work: keep the context files updated. An agent running on stale data makes stale recommendations. Monthly context file reviews keep the agents useful.
If you're building this kind of AI-powered workflow for a small business, The $20 Dollar Agency covers the full playbook from tool selection to cost management to the specific agent configurations that pay for themselves. Search "The $20 Dollar Agency" on Amazon Kindle.
Related reading
- A Markdown file is the best memory layer for your AI coding tool — how to structure the context files that power these agents
- Top AI CLIs and how to use them with our generators — the CLI tools for running agent roles
- How Claude cowork sessions actually work — the session patterns behind the researcher and analyst roles
- Two CLIs, one workflow: Codex alongside Claude Code — running multiple agent roles across tools
- How to validate an AI coding model before you trust it — verify the model before trusting it with financial or strategic work
This post is informational, not business or financial advice. Mentions of Anthropic, OpenAI, and all tools referenced are nominative fair use. No affiliation is implied.
