QUICK FACTS
● Launched: March 4, 2026
● Agent slots: Up to 4 custom agents per account
● Instruction limit: 4,000 characters per agent
● Setup path: Profile icon → Settings → Customize → Create Agent
● Free plan: Basic agent access. SuperGrok ($30/month) gets priority models and potential for more slots
● Invoke by: Name ("Harper, help me with...") or auto-engages by context
What Grok Custom Agents Actually Are
Grok Custom Agents are named AI personas you create and save to your account. Each agent has its own name, personality, instruction set, and focus area. When you start a new conversation, you pick which agent to talk to — or invoke one by name mid-conversation. Instead of talking to a generic Grok every time, you are talking to a specialist you configured.
The four default agents that ship with Grok 4.20 are Grok, Harper, Benjamin, and Lucas — four sub-agents that internally debate responses before surfacing a single answer in Heavy/Experts mode. The Custom Agents feature lets you create up to 4 of your own agents in the same slots, replacing the defaults with your specialisations. The architecture mirrors Grok's internal 4-agent system.
Custom Agents vs Grok Skills: Skills (launched May 18, 2026) are persistent instruction bundles for specific tasks, invoked with /commands. Agents are full AI personas that carry across entire conversations. Use Skills for repeatable task workflows; use Agents for a persistent working relationship with a specialist persona. Most power users use both.
Step-by-Step: Creating Your First Custom Agent
1 Sign in to grok.com
Sign in with your X account, Google, email, or Apple ID. Basic Custom Agent access is available on free accounts. SuperGrok ($30/month) or X Premium+ ($40/month) gives priority model access and higher rate limits.
2 Navigate to Settings
Click your profile icon in the bottom-left corner. Select Settings from the menu.
3 Open Customize → Create Agent
Inside Settings, select Customize. You will see your current agent slots (up to 4). Click Create Agent or Edit on an existing slot.
4 Name your agent
Choose a short, memorable name. You will invoke this agent by name in conversation (e.g. "Alex, review this code"). Avoid generic names — specific names help Grok activate the right agent by context.
5 Write the instructions (4,000 char limit)
The instruction field is your agent's system prompt. This is the most important step. See the instruction writing guide below.
6 Save and test
Save the agent. Start a new conversation. Invoke by name: "[Agent name], [your request]." Test across 5-10 different prompts to verify the instructions are working as intended before relying on the agent in production.
How to Write Effective Agent Instructions (4,000 Character Limit)
The 4,000 character instruction limit (reduced from 12,000 in late 2025 when Custom Agents launched) requires precision. Use this structure:
Agent instruction template
You are [NAME], a [specific role] specialising in [domain].
ALWAYS:
- [Specific behaviour 1]
- [Specific behaviour 2]
- [Tool usage: "always use x_search for X" / "always search web for Y"]
- [Output format: "always structure as [format]"]
NEVER:
- [Specific prohibition 1 — negative rules are highly effective]
- [Never use X type of language]
- [Never skip the step of Y]
TONE: [One sentence describing communication style]
OUTPUT FORMAT:
[Exact format you want every response in]
CONTEXT: [Any persistent context this agent needs — your role, company, audience]
The most important rule: Be specific and negative. "Be professional" is weak. "Never use emoji. Always use numbered headings. Always cite sources with [Source: URL]" is strong. Negative constraints are surprisingly effective at preventing the personality drift that affects ~30% of extended Grok conversations.
The 4-Slot Strategy — How to Allocate Your Agents
With only 4 agent slots, you need to be intentional. The most effective allocation covers the four core types of work most users do:
Slot 1: Research specialist
Uses x_search and web_search in parallel. Always cites sources. Produces structured intelligence briefs. Use for competitive research, market analysis, background preparation.
Slot 2: Writing specialist
Knows your tone, audience, and style. Never uses filler phrases. Always produces content in your specific format. Use for all content creation — posts, emails, documents, scripts.
Slot 3: Domain specialist
Deeply configured for your specific field — coding, finance, legal, marketing, healthcare. Knows your context, your industry's terminology, and the output formats your domain requires.
Slot 4: Real-time intelligence agent
Always uses x_search first. Always includes a timestamp. Always flags emerging signals. Use for monitoring, trend detection, news briefings — any time the live X data advantage matters most.
Preventing Personality Drift
Personality drift — where agents gradually revert to default Grok behaviour in long conversations — affects approximately 30% of extended chats. Four proven countermeasures:
Anchor prompt: Start each conversation with "You are [agent name] mode." A single reinforce line resets the agent context.
Use negative rules: "Never use emoji" and "Never use casual language" are more drift-resistant than positive instructions like "be professional."
Start fresh after 20 turns: Long conversation histories dilute instruction weight. Start a new conversation and reinvoke the agent rather than pushing past 20+ turns.
Repeat key rules: Include your most critical constraints twice in the instructions — once at the start and once at the end of the 4,000-character block.
Ready-to-use agent instruction templates: Grok Agent Instructions Examples — 8 complete Skills templates. Agent prompts library: Grok Agent Library — 50 ready-to-use agents. Grok vs ChatGPT agents: Grok Agents vs ChatGPT GPTs comparison. All AI news: June 2026 calendar.