Agents & Frameworks

The agentic layer of AI is where the interesting engineering problems live. This cluster covers the full stack of autonomous AI systems: the orchestration frameworks that sequence tool calls, the memory architectures that let agents stay coherent across sessions, the trust and autonomy models that determine how much rope you give a running agent, and the emerging protocol standards that connect those agents to external systems.

On the protocol front, the Model Context Protocol has become the connective tissue of the modern AI stack. Groundy has tracked MCP from its initial Anthropic release through GitHub’s registry play, the Zed and JetBrains integrations, and the ACP (Agent Communication Protocol) effort that targets multi-agent coordination at the registry level. The MCP registry now hosts thousands of servers; the question has shifted from “will this standard win?” to “how do you audit what you’re connecting to?”

Framework choices are less settled. PydanticAI and LangChain represent genuinely different philosophies—type-safe structured outputs versus flexible chain composition—and the right answer depends heavily on whether you’re building a one-shot pipeline or a long-running autonomous process. CrewAI, AutoGen, and Superpowers sit further up the abstraction stack, enforcing workflow discipline on top of bare tool-use APIs.

Memory architecture is the underappreciated dimension. Hindsight, vector stores, and session management each solve a different failure mode—context overflow, retrieval hallucination, and cross-session amnesia are distinct problems with distinct solutions. The pattern of conflating them is one reason production agent deployments underperform their demos.

Groundy covers this cluster with an engineering orientation: architecture decisions, production failure modes, benchmark methodology, and the governance questions—autonomy bounds, human-in-the-loop thresholds—that don’t show up in vendor benchmarks but matter most when agents run unattended.

The autonomy question is not just philosophical. Agents that can browse the web, execute code, push commits, and make API calls on your behalf are also attack surfaces. Prompt injection through tool outputs, malicious content retrieved by browsing agents, and supply-chain attacks on MCP server packages are all documented. This cluster covers the security dimensions of agent architecture alongside the engineering ones—because in 2026 you cannot separate them.