AI and ML
The AI-powered workforce: Cultivating the new breed of talent for 21st century operations
_web_limited%20text%20banner_1800%20x%20300:360?wid=360&hei=300&qlt=90%2C0&resMode=bicub&op_usm=1.75%2C0.3%2C2%2C0)
The Model Context Protocol (MCP) introduces a semantic context layer that allows an AI agent to seamlessly access meaningful context (tools, memory and state), driving more specific and responsible AI output.
Unlike traditional APIs focused on function execution, MCP emphasizes context-sharing, enabling more accurate, grounded and cooperative AI behavior across tools — with significantly reduced friction in interoperability.
MCP also makes it easier to integrate and implement retrieval-augmented generation (RAG), helping AI agents to fetch details or resources dynamically as needed. This mitigates the limitations of LLMs’ context windows and allows agents to remain focused on the task at hand while retrieving relevant data on demand.
MCP vs APIs: Context vs. function
Deciding whether to implement MCP starts with a fundamental architectural question:
“Does context matter here?”
Traditional APIs are great for executing repeatable functionalities. MCP, however, is for systems that learn, reason and collaborate — systems where understanding what just happened matters as much as what to do next.
Take AI-assisted software development, for example. Coding agents need to understand business logic, architectural constraints, tech debt and user feedback. This has created significant demand for standardized context-as-a-service; MCP is emerging as the preferred approach to supplying that context across tools like IDEs, AI assistants and coding models.
The table below offers a comparison of APIs and MCP:
Comparing APIs and MCP
Feature |
API |
MCP |
Core function |
Provides function calls. |
Provides semantic context. |
Response behavior |
Fixed input-output driving rule-based actions. |
Dynamic response based on context driving intelligent decisions. |
Target consumer |
Software applications. |
AI models or agents. |
Flexibility |
Functional, standardized communication. |
Semantic, context-aware conversations. |
Example |
getUser(id) |
“This user expressed dissatisfaction in the last conversation.” |
SRE: A textbook use case for MCP
Site Reliability Engineering (SRE) is one of the clearest use cases for MCP. SRE workflows require:
Deep situational awareness.
Multi-agent collaboration.
Real-time decision-making across disparate systems.
As AI becomes more embedded in reliability practices, leaders like Rootly and Chronosphere are already building MCP-compatible capabilities into their incident and observability stacks.
Below is a comparison of the difference between using and not using MCP in an SRE context:
| Job | Without MCP | With MCP |
| Alert triggers | Chronosphere generates alert -> Rootly creates incident -> info lost across tools. | Alert context packaged as IncidentContext object -> consumed directly by RCA agent. |
| Root cause agent | Must be re-prompted with full alert, logs, context manual. | Dynamically request further alert or incident details. |
| Action planning agent | Needs repeated background explanation, user input for multi-step reasoning. | Leverages shared context and enables seamless multi-step tool use with LLM-friendly inputs and outputs. |
| Resolution development | Context-switching: engineers copy/paste logs and context into IDE prompts. | Agent pulls from IncidentContext, codebase metadata and runtime context directly in IDE. |
| Report generation | Hard to align with what actually happened. | Uses same context used by prior agents -> coherent, accurate summary. |
Key MCP-powered use cases in SRE
MCP is enabling a new level of intelligence and coordination across the SRE ecosystem. Here are three core use cases where it's driving real impact:
1. Context-aware observability engineering
AI agents can help correlate alerts, detect anomaly clusters and align issues with service topologies.
The context here includes SLOs (service-level objectives), historical trends, detailed telemetry and log data, ownership, past alerts and incidents.
The MCP host will be an observability platform, such as Chronosphere.
MCP clients include things like RCA agents, alert deduplication bots and FinOps optimizers.
2. AI-assisted incident investigation and triage
Agents can carry forward rich semantic context as they investigate causes, propose fixes or escalate incidents.
The context here includes alert state, system health, past incidents, prior remediations, who solved the incident, real-time log context.
The MCP host could be an incident management platform (like Rootly MCP).
MCP Clients here might be Slack bots, summarization agents, ticket generators, RCA agents, incident investigation and resolution agents.
(If you want to see how this works in practice, check out the Rootly MCP Server Demo.)
3. Semantic handoffs across systems
MCP ensures AI agents working across Slack, Jira, Confluence or runbooks can hand off not just statuses, but shared understanding too.
The context here includes things like the incident timeline, related tickets and issues, standard operating procedure (SOP), recent updates, feature details and past response actions.
The MCP Host could be a knowledge management platform (e.g. MCP Atlassian).
The MCP clients include workflow orchestrators, playbook automation runners, content creators and human-in-the-loop assist tools.
Why MCP matters for SRE and platform leaders
If you lead SRE, platform engineering or intelligent operations, it's vital to understand that MCP isn't about replacing APIs: it’s about augmenting them. MCP introduces a dynamic context layer that lets AI agents act with awareness, use the right tools, collaborate with memory, generate more accurate outputs and evolve their behavior over time.
It’s a shift from “API calls” to “contextual reasoning.” And that shift is foundational to building autonomous, intelligent reliability systems.
Disclaimer: The statements and opinions expressed in this article are those of the author(s) and do not necessarily reflect the positions of Thoughtworks.
