Your Agent Framework Is a Chassis, Not a Convenience

An LLM call is stateless and amnesiac. It forgets everything the instant it finishes. Turning that single turn into a coherent multi-step task, deciding when to call a tool, what to do with the result, when to stop, is a separate job. That job is the framework, and skipping it does not remove the work. It just moves the work onto whoever has to debug the mess later.

The Details:

• Teams that avoid adopting a framework almost always build one anyway, by accident. A happy-path LLM call gets patched with string parsing for tool names, a try/except for bad JSON, a hand-rolled retry for wrong tool picks. Six weeks in, there are hundreds of lines of undocumented glue code that only one person understands.

• Three failure modes show up without exception. State mismanagement, where teams either cram everything into context until quality craters or prune so hard the thread disappears. Missing observability, so when step fourteen of twenty goes wrong nobody can reconstruct what the model saw or decided. And an unspoken assumption that the demo path is the only path, when real runs fail in ways demos never do.

• The market sits closer to the early Perl-and-CGI moment than to a Rails-style consolidation. Models keep absorbing capability that today requires scaffolding, so teams that build thick custom layers on top of one framework often find, six months later, that half the scaffolding is obsolete. The safer move is staying loose enough to swap parts, not betting on a winner.

• Four habits outlast any framework swap: scope every tool like a permission grant, push irreversible actions like merges or refunds behind a human checkpoint, log enough to reconstruct any decision after the fact, and write tool contracts precisely, since sloppy interfaces produce sloppy agent behavior no matter how good the model underneath is.

Bottom Line: The framework you pick this year will likely be gone in three; the discipline of scoping, reversibility, and observability will not.