Topic
Design and internals of frameworks that run LLM-driven agents, including runtime loops, tool calling and framework comparisons.
AutoGen has about 60,400 GitHub stars, the most of three leading agent frameworks, but its own README puts it in maintenance mode. A dev.to audit verified on 2026-08-16 says to weigh installs and release history before stars when starting a new project.
Reality
Strands and CrewAI agents exited 0 on all six runs where a broken verification tool checked nothing, a recorded-proxy test on dev.to found. Only tool-call traffic captured outside the framework showed the failures.
Reality
Only a renamed tool argument got through Strands, LangGraph and CrewAI cleanly in a 36-run schema-change test posted on dev.to. Harsher changes let Strands and CrewAI exit 0 with nothing verified while LangGraph crashed outright, so each framework needs its own schema-change test.
Reality
In apowerb the module the runtime imports is generated and disposable, and every instruction, tool and model string is read out of a database row at load time. The cost is a filesystem holding derived state.
Reality
Fluid compute caps Hobby functions at five minutes and Pro at 800 seconds, so an agent working a long task list needs a queue or a different host. The cheap VPS escape hatches have repriced too.
Reality
The MIT-licensed TypeScript framework at v0.16 exposes seven named run phases with hooks on each side, and its case for harness over model rests on one incident-triage run that a 4B local model and Claude both finished.
Reality
One extension method turns an IChatClient into a bounded agent loop with tool approval and history persistence already wired. The default exit condition is an exact marker string the model has to emit.
Reality
Arcjet's new product evaluates an agent's next action inside application code, returning an allow or deny before the tool call runs, with policies written in Rego and managed centrally by the security team.
Reality
Microsoft's proof of concept launched calc.exe from a hotel-finder agent, and a second CVE in the .NET SDK let a prompt drop a file into the Windows Startup folder. Both fixes shipped the same day.
Reality
The durable output of OpenAI's new agent cookbook is an eval suite generated from human and model feedback on five runs of one fictional company, plus a handoff file that tells Codex what to change next.
Reality
The harness lost its hidden system prompt, 43% of its builtin tool descriptions and its todo list middleware. LangChain's own footnote says reward confidence intervals span zero for every model tested, so the evals settle the token saving more firmly than the quality.
Reality
A dev.to teardown describes a framework where the messaging adapter is the frame and the agent loop lives inside it. The session lane is where you pay for that choice.
Reality