Introducing Custom Agents: Build Your Own AI Specialists in Minutes

Create specialized AI agents that understand your pipelines, your context, and your workflows

Cody

Cody Peterson

cody.peterson@ascend.io

I spend a lot of time talking with data engineering teams, and the same pain points come up in every conversation: pipelines fail at the worst possible moments, incident response eats up way too much time, and every team has unique operational needs that no generic tool can anticipate.

Most teams end up building custom scripts, setting up basic alerting, and hoping for the best. But there's always been this gap between what generic automation can do and what data teams actually need—automation that understands their specific pipelines, business context, and operational procedures.

Today, I'm excited to share how we're closing that gap with Custom Agents, a new capability that lets data teams build intelligent, context-aware automation in minutes rather than months.

The Problem: Automation that doesn't understand context

Data pipeline automation today feels like it's stuck in the past. When something breaks, you might get an alert, but then you're back to manual work: digging through logs, understanding downstream impact, coordinating fixes across teams, updating stakeholders.

The fundamental issue is that traditional automation operates in isolation from your data platform. It can't answer questions like "What changed in the recent deployment that might have caused this?" or "Which business units will be affected if this pipeline stays down?"

Even when teams invest weeks building custom automation, these tools lack the context to make intelligent decisions. They're essentially fancy if-then statements that execute the same response regardless of the situation.

A Different Approach: AI that sees what you see

At Ascend, we've been working on something we call Agentic Data Engineering—the idea that AI agents should be full participants in data operations, not just code assistants bolted onto existing tools.

Our AI assistant Otto already helps teams with routine tasks like generating documentation, suggesting code improvements, and monitoring pipeline health. But we kept hearing from teams: "This is great, but we have unique operational needs that go beyond what any built-in system can handle."

That's where Custom Agents come in. They're specialized versions of Otto that you can build for specific use cases, configured with your organization's context and connected to your existing tools.

‍

SEE IT IN ACTION: Check out Agentic Data Engineering

‍

What makes custom agents different

The key insight is that effective automation requires the same context that human engineers rely on. When an experienced data engineer investigates a pipeline failure, they don't just look at error logs—they consider recent code changes, understand downstream dependencies, know which teams to notify, and can recognize patterns from similar past incidents.

Custom Agents work the same way. They have access to comprehensive metadata about your data pipelines: execution history, dependency relationships, recent changes, performance patterns, and business context. This unified view enables them to make informed decisions rather than executing rigid scripts.

Take the example of a pipeline failure. Here's how it works in practice—when a production pipeline fails, a Custom Agent can automatically:

Analyze the failure using execution logs and metadata
Compare against similar historical incidents to identify likely causes
Assess downstream impact based on data lineage
Create detailed incident reports with suggested remediation steps
Notify the right teams through your existing communication channels
Track the incident through to resolution

All of this happens automatically, with the same level of analysis you'd expect from an experienced engineer who knows your systems inside and out.

Users can configure agentic automations to trigger on events within their system including pipeline successes and failures.

Building intelligence that understands your business

What sets Custom Agents apart from traditional automation is how easy they are to build and customize. Instead of weeks of development, you define agent behavior using simple markdown files with natural language instructions.

Here's what a "Production Incident Agent" looks like:

‍

Production Incident Response Agent

---
otto:
  agent:
    name: Production Incident Response
    mcp_servers:
      - slack
      - github
    model: gpt-4o
    model_settings:
      temperature: 0.1
---

# Production Incident Response Agent

You are a senior data engineer specialized in production incident response and team coordination.

## Your Role
When a production pipeline fails:

1. **Analyze the failure** using execution logs and recent system changes
2. **Assess business impact** based on downstream dependencies and SLAs  
3. **Create detailed incident report** with root cause analysis and remediation steps
4. **Notify stakeholders** through appropriate channels with context and urgency
5. **Track incident to resolution** and update all relevant parties

## Escalation Rules
- **Critical failures** (customer-facing data): Immediately notify #critical-alerts and page on-call engineer
- **Standard failures** (internal dashboards): Post to #data-ops with analysis and timeline
- **Performance issues** (slow but functional): Create GitHub issue and notify team lead

## Communication Guidelines
- Include pipeline name, environment, and business impact in all notifications
- Provide links to relevant logs, monitoring dashboards, and incident tickets
- Use clear, non-technical language when notifying business stakeholders
- Update all parties when incident is resolved with summary of fix

Always include specific remediation steps based on similar past incidents and current system context.

‍

This agent understands your team's specific escalation procedures, knows which stakeholders to notify for different types of failures, and can automatically create tickets in your project management system with the right priority and context.

Or you might build a "Code Review Assistant" that enforces your team's standards:

‍

Code Review Assistant Agent

---
otto:
  agent:
    name: Code Review Assistant
    model: gpt-4o
    tools: ["*"]
---

# Senior Code Review Assistant

You are a senior data engineer who reviews pipeline code for quality, performance, and maintainability.

## Review Criteria

**Code Quality**
- Proper error handling for all data operations
- Clear variable names that reflect business concepts
- Adequate logging for debugging and monitoring

**Performance** 
- Efficient data processing patterns
- Appropriate partitioning strategies
- Resource usage optimization

**Maintainability**
- Clear documentation of business logic
- Modular design that supports testing
- Consistent naming conventions

## Review Process
1. Analyze the code changes in context of the broader pipeline architecture
2. Check for adherence to team coding standards
3. Identify potential performance bottlenecks or error conditions  
4. Suggest specific improvements with code examples
5. Highlight any security concerns or data handling issues

Provide constructive, specific feedback that helps engineers improve their skills while maintaining high code quality standards.

‍

The agents connect to your existing tools—Slack, GitHub, PagerDuty, Jira—through standard integrations, so they enhance your current workflows rather than requiring you to change tools.

Beyond Incident Response: Proactive operations

While incident response gets a lot of attention, Custom Agents excel at proactive operations too. Teams are building agents that:

Enforce Governance: Automatically review code changes for compliance with security policies and coding standards
Optimize Performance: Monitor resource usage patterns and suggest optimizations based on cost and performance targets
Streamline Communication: Generate executive summaries of pipeline health and automatically distribute reports to stakeholders

The common thread is that these agents understand your specific business context, operational procedures, and team dynamics.

Making advanced AI accessible

One of our core beliefs is that powerful AI capabilities shouldn't be limited to teams with extensive machine learning expertise. Custom Agents democratize sophisticated automation by providing a simple interface backed by advanced AI.

As our CEO Sean Knapp puts it: "Building powerful and safe AI agents shouldn't take weeks. Teams should be able to build, test, and deploy sophisticated automation safely in just a couple of minutes."

This accessibility is crucial because every data team has unique needs. The specific failure patterns you see, the stakeholders you need to coordinate with, the business logic that drives your decisions—these details matter, and they're different for every organization.

The broader shift toward Agentic Data Engineering

Custom Agents represent part of a broader shift we're seeing in data operations. Instead of expecting humans to manually manage every aspect of pipeline operations, we're moving toward systems where intelligent agents handle routine tasks and escalate when human expertise is truly needed.

This isn't about replacing data engineers—it's about amplifying their capabilities so they can focus on strategic work rather than operational toil. The best data engineers spend their time solving complex business problems, not babysitting pipelines or manually coordinating incident response.

What's next

We're excited to see what teams build with Custom Agents. Every data engineering team has unique operational challenges, and we expect to learn a lot about use cases we haven't even considered yet.

The early feedback has been encouraging. Teams report faster incident response times. But most importantly, engineers are spending more time on the creative, strategic work that makes a real business impact.

‍