
Agno 3.2.1, released November 17 and reviewed for this edition, makes a session’s compaction records readable through AgentOS. It also removes a default 1024-token output cap from its OpenRouter, Perplexity and Requesty integrations when callers leave the limit unset.
The two changes concern different limits. Compaction changes the context retained for later work; the provider change affects how much output the client requests. Neither makes a model’s native input capacity unlimited.
What the compaction endpoint returns
The compaction change lists records newest first, including message counts, before-and-after token counts, creation time and an associated run where applicable. It does not return the summary text or archived messages.
The source describes stored records for SQLite and PostgreSQL. A backend that does not store them can return an empty list, so an empty response is not by itself proof that no compaction ever happened. The tagged route also avoids asking a remote AgentOS that may predate the endpoint.
Our analysis: a smaller token count does not prove preserved meaning
Persistent records help explain why the context changed between turns and after a page reload. That is useful operational evidence. It does not establish which details a summary retained or whether a later answer still had the evidence it needed.
A meaningful quality check would compare subsequent work against the information that mattered before compaction. The right question is not only how many tokens disappeared, but whether an instruction, exception or unresolved uncertainty disappeared with them.
That also explains why returning counts without summary content is a limited interface. It supports a history view; it is not a complete audit of the model’s retained information. Applications should describe that distinction clearly to users.
The output-cap fix is narrower than unlimited generation
The provider patch replaces the integrations’ default 1024 value with an unset limit. The request then leaves the upstream model or service to apply its own constraints. Explicit caller-supplied limits remain forwarded.
Removing a client-imposed ceiling can prevent an avoidable truncation, but it cannot promise that every response will complete. Model limits, provider policies and the workload still apply. We did not measure response length, cost or retention quality.
Our Agno repository profile links the Apache-2.0 source. This is source-based reporting, not an installed integration in our launcher. No model weights, provider credentials or local runtime were changed.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- agno-agi/agno ↗View project details
Primary-source review and original editorial analysis. No software installation, model inference, paid service call, security validation or independent benchmark performed.

