Frameworks

Frameworks are the runtimes teams install to make agents: graphs of steps, role crews, tool binders, and memory hooks. They sit at the center of Agentic. A framework is not Serving (weights/inference) and not RAG alone (though many embed RAG helpers). It is the control loop: what runs next, which tool fires, when a human gate is required.

Technically compare orchestration models (state graphs vs conversational crews vs DSPy-style programs), tool schemas, streaming, checkpointing, and how hard it is to keep side effects attributable to a case. Failure modes: demos that cannot recover, unbounded tool spend, and frameworks that silently become the system of record because nobody wired a ledger.

This cell is crowded (twenty-plus packs). Representative cuts:

• LangGraph-adjacent and Crew AI / Autogen / AG2 style multi-agent composition for role-shaped work.
• DSPy for programmatic LLM pipelines with optimizers rather than free-form chat agents.
• Google ADK, Agno, CAMEL, AgentScope, and peers compete on DX, cloud bindings, and multi-agent patterns.
• Thinner kits (Atomic Agents, BeeAI Framework, …) trade ecosystem breadth for smaller surfaces.

Prefer a framework when you need a durable control loop and tool policy; prefer Workflow Orchestration when the bottleneck is long-running business workflows with timers/retries; prefer Coding Agents when the environment is a repo. Feature maps such as Agent Lifecycle Management and Agent Building Platform mark the product job: build and run agents without reinventing the loop each time.

Keep Monday out of the runtime. Frameworks propose and execute under gates; boards remain read-models. If the framework’s state store is the only place truth lives, the architecture has already failed the observe→ledger path.

Compare frameworks on recovery and attribution, not on logo count. Checkpointing, human-in-the-loop nodes, and typed tool errors determine whether a failed estimate run can resume or must restart. Prefer graph-shaped runtimes when the business process has named stages; prefer crew/role frameworks when the work is parallel specialists; prefer compiler/optimizer styles (DSPy) when the pipeline is a program you want to tune against metrics. Keep secrets and customer PII out of framework default logs. The framework is connective tissue — judgment stays at the gate.

For this research catalog the acceptance test is boring and strict: after wiring the pack, can a second person replay the same scenario and get the same measurable outcome? If not, the page is still a brochure. Document the scenario, the expected number, and the log line that proves it. That habit is what separates a deepened taxonomy cell from a landscape screenshot.

Also record what you refused: tools not connected, corpora not ingested, metrics not collected. Refusals keep the cell honest when marketing pages claim universality. A taxonomy essay that only lists winners teaches less than one that names the skip conditions next to the install path.