4 workflows to adopt AI agents, beyond code generation
Generative AI has transformed software development so dramatically that you can spin up entire services in hours, yet understanding what went wrong with those services still demands painstaking work across fragmented tools. From code generation to code review, coding agents handle the build side. But production debugging? That's still manual. Take the following example:
| Coding | Production |
|---|---|
| To write a service: You open an AI-native development environment and ask AI to "Create a payment service that handles retries and timeouts". AI generates implementation with error handling, using context of your code base | When the same service is experiencing high latency: You start with a hypothesis → Check Datadog for metrics → Switch to Loki for logs → Cross-reference deployment history → Correlate timestamps →... And so on |
The problem isn't AI's capability; it's how we architect AI systems. Most engineering teams still use AI-powered tools to execute the same workflows faster, not reimagining how software development and production operations should work end-to-end.
At Resolve AI, we've been building multi-agent systems for engineers to work on production systems. We've been advocating that engineering should be AI-native (where engineers primarily interface with autonomous agents to work on production systems), while most gen AI conversations in software engineering were centered on writing AI generated code with copilots and coding assistants.
We recently presented our approach to Stanford's graduate AI program, diving deep into the AI agents and their architectural patterns that enable AI-native engineering workflows.
AI-Native Engineering is where engineers primarily interface with AI to orchestrate their work: be it writing code or working on production systems. AI-native is a significant departure from just “using AI” where engineers are still interfacing with their systems and tools, but using AI to speed up individual steps of the process.
Here is an example workflow to showcase the distinction.
AI-Assisted: You use AI tools to work faster on complex tasks. The workflow remains human-centric: Engineer → Systems and tools → Correlation → Action. Engineers still interface with tools, just using AI to perform individual tasks faster
AI-Native: AI becomes your primary interface for production work. The workflow becomes AI-led: Engineer → Natural Language Request → AI System → Response / Action. Engineers set goals and let AI agents handle the operational work
Take incident response as an example. In AI-assisted workflows, you're still generating hypotheses, deciding which evidence matters, and manually correlating signals across tools. AI helps with data retrieval and analysis, but you're doing the heavy lifting in investigation.
AI-native incident response operates differently: AI agents perform real-time triage of investigation priorities, generate competing hypotheses in parallel, and refine theories through successive iterations based on cross-system evidence. Instead of asking "Can you analyze these logs?" you say "Resolve this checkout failure" and agents coordinate the entire investigation.
This isn't just faster. It changes what problems deserve engineering attention. When AI agents handle log analysis, metric correlation, and deployment timeline reconstruction, engineers operate at a higher-level, focusing on architectural decision-making and system design rather than tactical investigation.
The shift requires persistent AI agents, not just AI tools. While AI models like those from OpenAI or Anthropic can accelerate individual tasks, only stateful agents can maintain investigation context, coordinate across multiple tools, and execute complex tasks across the full incident lifecycle autonomously.
Modern production systems exhibit what academics call "irreducible interdependence": understanding them requires specialized knowledge across domains that cannot be unified into a single coherent model. This is the insight most builders miss: No single AI tool or set of AI models can maintain expert-level knowledge across all these domains while coordinating a real-time investigation.
For example: When API latency spikes 10x during a critical incident, the investigation requires simultaneous specialized agents performing real-time analysis: correlating traces across 50+ microservices, analyzing slow database queries and connection pool exhaustion, checking recent deployments and infrastructure changes, scanning auth logs for security anomalies, evaluating auto-scaling functions against current load patterns, and analyzing support tickets for customer impact with SLA context. Each of these functions requires domain-specific expertise and contextual data that no single system could effectively maintain.
At large-scale, as system complexity increases, individual AI tools lack the adaptability to handle exponential growth in context requirements. This is where multi-agent systems can scale by combining orchestration and individual domain specialization. This matrix provides an overview for engineering leaders. Find the row that matches your current state to understand its limitations:
This framework reveals the technical limitations at each level:
| Approach | What it is | Where the approach breaks | Cause for limitation |
|---|---|---|---|
| LLM | Use LLMs (such as ChatGPT) for individual tasks like explanations, analysis, and documentation | Engineers still do majority of the operational workload | Single-pass generation prone to hallucinations, with no feedback loops or real-world integration |
| LLM + Tools | AI can call functions to fetch data from monitoring systems on command | Cognitive workload of correlation remains on humans | Limited context windows, no persistent state management across tool interactions |
| Single Agent | AI follows investigation workflows independently | Sequential investigation, gets stuck on wrong hypotheses without validation | Cannot manage diverse reasoning strategies or parallel investigation paths |
| Multi-Agent | Through multi-agent collaboration, specialized AI agents coordinate parallel investigations and combine their output into a unified diagnosis | Requires investment in coordination protocols | Distributed intelligence needs formal communication schemas and conflict resolution |
The progression reveals a fundamental architectural truth: Each level hits a different scalability ceiling. LLMs lack a persistent state. Tool-augmented LLMs can't maintain investigation context across multiple chats. Even with sophisticated prompt engineering, single agents become decision-making bottlenecks as system complexity grows. Only multi-agent systems can break through the sequential reasoning constraint that limits all previous approaches. They enable parallel hypothesis testing while single agents must investigate sequentially, making them fundamentally unsuitable for the temporal demands of production incidents.
No off-the-shelf agent framework like LangChain solves this alone. Building production-ready multi-agent systems requires a rare combination of deep domain expertise and AI engineering prowess. Most attempts fail because teams have expertise in one area but not both. Here’s why this dual expertise is needed:
At Resolve AI, our team includes engineers with over two decades of experience in production systems, founders who co-created OpenTelemetry (one of the most impactful open-source observability projects in the ecosystem) and researchers with deep artificial intelligence expertise who are the minds behind Google DeepResearch and Gemini Agents. This combination lets us build systems that don't just understand that "payment failures are bad" but know to check connection pool metrics and correlate with upstream service degradation. All the while managing complex agent orchestration that prevents circular investigations and maintains coherent narrative threads across parallel execution paths.
Resolve AI is your always-on AI SRE that helps you resolve incidents and run production. With Resolve AI customers like Salesforce, Zscaler, and Coinbase have increased engineering velocity and systems reliability by putting machines on-call for humans and letting engineers just code. Learn more about AI-native engineering workflows at resolve.ai.
Content
Spiros Xanthos
Founder and CEO
Spiros is the Founder and CEO of Resolve AI. He loves learning from customers and building. He helped create OpenTelemetry and started Log Insight (acquired by VMware) and Omnition (acquired by Splunk), most recently he was an SVP and the GM of the Observability business at Splunk.
Gabor Angeli
Research Engineer
Gabor Angeli brings extensive AI expertise, most recently at Google DeepMind and Square. His work on products like Gemini and Square Assistant touches millions of users daily. He joined Resolve AI to build Agentic AI systems that help engineers understand and navigate production systems.
Bharat Khandelwal
Research Engineer
@ Resolve AI
Bharat is a Research Engineer at Resolve AI, where he builds agentic systems that enable large language models to debug and operate production software infrastructure. Prior to Resolve, he led machine learning initiatives at WorldQuant, designing transformer-based architectures for macroeconomic forecasting and incorporating LLM-driven sentiment signals from unstructured data. He has also worked at Moveworks on enterprise NLP systems and at Tower Research Capital, where he developed low-latency ML strategies for high-frequency trading. Bharat holds an M.S. in Computer Science from Stanford University, with a specialization in Artificial Intelligence, and a B.Tech. (Honors) in Computer Science from IIT Bombay.
Hear AI strategies and approaches from engineering leaders at FinServ companies including Affirm, MSCI, and SoFi.
Learn how to evaluate an AI SRE to ensure they run in your unique production environments. This guide explains the five agentic pillars, six key evaluation dimensions, and enterprise readiness criteria that separate production-ready AI SRE from experiments.
100 Engineering software engineering executives joined Resolve AI and other luminary leaders to discuss the accelerated evolution of agentic AI in software engineering from coding to managing production systems.
