HalosHalosDocs

Training Your Agents

Training is what transforms a generic AI into your AI. A freshly deployed agent knows its specialty and has a default personality, but training teaches it your specific knowledge, rules, and communication style.

How Training Works

When you add training data to an agent, that data gets injected into the agent's system prompt every time it responds. The agent doesn't just memorize facts — it uses your training data as context to shape every response.

This means:

  • An email agent trained on your company's support docs will write replies using your terminology
  • An agent with instructions to "always respond in bullet points" will do exactly that
  • An agent trained with example conversations will mirror the tone and style you demonstrate

Training takes effect immediately. The next message you send will use the new context.

Three Types of Training Data

Knowledge

Facts, documents, reference material, or any text your agent should know about.

Use it for:

  • Company documentation, FAQs, or wikis
  • Product specs and feature lists
  • Personal notes and preferences
  • Research papers or article summaries
  • Client profiles and project details

Real-world example — training an email agent:

Title: Email response guidelines

Content: When responding to customer emails: always greet by first name, acknowledge their specific issue before offering solutions, include relevant KB article links when available, and sign off with "Best, [my name]". For billing issues, CC finance@company.com. For technical issues over 24 hours old, escalate to the engineering channel.

Instructions

Custom rules and behavioral guidelines that tell the agent how to behave.

Use it for:

  • Communication style rules
  • Guardrails and boundaries
  • Process rules and standard operating procedures
  • Role-specific behavior definitions

Real-world example — training a research agent:

Title: Research standards

Content: When researching a topic: always cite sources with URLs. Prioritize primary sources over blog posts. Flag when information is older than 6 months. Present findings as a structured brief with an executive summary (2-3 sentences), key findings (bullet points), and sources section. When I ask for "quick research," limit to 3-5 key points. When I ask for "deep dive," be comprehensive.

Examples

Example conversations showing the agent exactly how to respond in specific scenarios.

Use it for:

  • Demonstrating your preferred tone and style
  • Teaching domain-specific Q&A patterns
  • Showing how to handle edge cases
  • Few-shot learning for specific task types

Real-world example — training a code agent:

Title: Code review style

Content: User: Review this PR for me Assistant: Here's my review:

Summary: [one sentence on what the PR does]

Issues (must fix):

  • [critical problems]

Suggestions (nice to have):

  • [improvements]

Good stuff:

  • [what's done well]

Verdict: [Approve / Request changes / Needs discussion]

Adding Training Data

  1. Navigate to the agent's profile page (tap their avatar in The Yard)
  2. Scroll to the Training section and tap Train [agent name]
  3. Click Add Training Data
  4. Select the type (Knowledge, Instruction, or Example)
  5. Add a title and the content
  6. Click Add

Training by Use Case

Here are detailed training recipes for common agent types:

Email Agent

Goal: An agent that triages your inbox, drafts replies in your voice, and knows your priorities.

| Type | Title | What to Include | |------|-------|----------------| | Knowledge | Key contacts | Names, roles, and context for people you email regularly | | Knowledge | Email templates | Templates you use for common replies (intro, follow-up, scheduling) | | Instruction | Triage rules | How to categorize emails: urgent vs. FYI vs. can-wait. What constitutes "urgent" for you | | Instruction | Tone guide | Your communication style — formal/casual, signature, greeting preferences | | Example | Reply examples | 2-3 examples of real emails you've sent (anonymized) showing your actual style |

Research Agent

Goal: An agent that finds information the way you want it, formatted the way you need it.

| Type | Title | What to Include | |------|-------|----------------| | Knowledge | Industry context | Your industry, key players, terminology, and what matters to you | | Knowledge | Preferred sources | Publications, databases, or websites you trust | | Instruction | Output format | How you want research presented (brief, report, bullet points, etc.) | | Instruction | Focus areas | Topics you care about, things to watch for, recurring research themes | | Example | Research output | An example of a research summary you'd consider "good" |

Code Agent

Goal: An agent that reviews code your way, understands your stack, and follows your conventions.

| Type | Title | What to Include | |------|-------|----------------| | Knowledge | Tech stack | Languages, frameworks, libraries, and tools your team uses | | Knowledge | Conventions | Coding standards, naming conventions, file structure patterns | | Instruction | Review priorities | What to focus on (security, performance, readability, etc.) | | Instruction | PR process | Your team's PR workflow, approval requirements, CI/CD notes | | Example | Code review | An example code review showing your team's feedback style |

Custom Agent (Personal Assistant)

Goal: A general-purpose agent that knows your life and handles whatever you throw at it.

| Type | Title | What to Include | |------|-------|----------------| | Knowledge | About me | Your role, company, interests, goals — the more, the better | | Knowledge | Key projects | Current projects, deadlines, and priorities | | Instruction | Communication | How you like information presented (concise vs. detailed, format preferences) | | Instruction | Boundaries | Topics to avoid, things you don't want help with | | Example | Task handling | An example of how you'd want a task broken down and handled |

Managing Training Data

On the training page, you can see all entries for an agent:

  • Toggle on/off — Temporarily disable an entry without deleting it. Useful for A/B testing different instructions.
  • Expand — Long entries are truncated by default. Click to see the full content.
  • Delete — Permanently remove an entry.

The stats at the top show how many knowledge entries, instructions, and examples the agent has.

Best Practices

Start Small, Then Iterate

Add a few key pieces of knowledge and one or two instructions. Chat with the agent to see how it behaves, then refine. It's much easier to iterate on a few entries than to debug 50 at once.

Be Specific

Instead of "be helpful," try "when the user asks about pricing, always mention the free tier first, then compare the Pro and Enterprise plans side by side with a comparison table."

Vague instructions produce vague results. Specific instructions produce consistent, useful behavior.

Use Examples for Tone

If you want a specific communication style, examples are the most effective teaching method. The agent will pattern-match against your examples more reliably than following abstract instructions.

Title Entries Clearly

Good titles make it easy to manage training data as it grows. Use descriptive names like "Q3 product changelog" or "Escalation handling rules" rather than "Info 1."

Disable Instead of Delete

If you're not sure whether an entry is helping, toggle it off instead of deleting it. Chat with the agent to see how it responds without that entry. You can always re-enable it.

Keep Knowledge Fresh

Review your training data periodically. Outdated information (last quarter's roadmap, a former team member's contact details) can confuse agents. Update or disable stale entries.

Training vs. Knowledge Base

Training data and the Knowledge Base both store information, but they serve different purposes:

| | Training Data | Knowledge Base | |---|--------------|---------------| | Scope | Per-agent | Account-wide | | Purpose | Shape a specific agent's behavior | Store facts about you that any agent can access | | Best for | Instructions, examples, agent-specific knowledge | Personal info, preferences, general context | | How it's used | Injected directly into the agent's system prompt | Searched contextually when relevant |

Use training data for agent-specific behavior. Use the knowledge base for general information about you that all agents benefit from.

Limits

Training data is included in the agent's context window, so there's a practical limit before responses start to degrade. Guidelines:

  • Knowledge: Up to ~20 entries works well. Keep individual entries focused.
  • Instructions: 5-10 well-written instructions beat 50 vague ones.
  • Examples: 3-5 high-quality examples per scenario are usually enough.

If you have more than this, prioritize the most impactful entries and consider moving reference material to the Knowledge Base instead.

Next Steps

Your agent is trained — now put it to work.