LoginSign Up
← Back to all sparks
A

Agno

DEVOPS
Velocity10.0

Agno's whole visible changelog is one release, v2.7.3, split across six rows.

agent-frameworkmcpsession-storagehuman-in-the-looptool-integrations
◆Current state
Every entry in this window is v2.7.3, dated 14 July 2026 — the feed publishes one row per change note rather than one per release, so a single version fills the timeline. The release adds ValkeyDb as an in-memory store for agent, team and workflow session state, RedmineTools for reading, creating, commenting on and logging time against issues, and wider human-in-the-loop confirmation, input and feedback support in AG-UI. The rest is MCP correctness: a clear error when a server is unreachable, and support for tools returning only structuredContent with no text block.
◆Where it's heading
The framework is filling in the operational layer rather than the agent abstraction. Session state gets a storage tier chosen for latency instead of durability, saving that state stopped making a model call it did not need, and MCP failures now surface legibly instead of as stack traces. Human-in-the-loop is being treated as a first-class interface concern in AG-UI rather than something callers build around. This is the work of a framework being run in production by people other than its authors.
◆Prediction
Expect more storage backends on the ValkeyDb pattern and further MCP response-shape handling, since both are being driven by what breaks in deployment rather than by a roadmap visible in these notes.

◆Recent moves

  1. 12d ago

    Give your agents a voice with Smallest AI

    SmallestTools adds text-to-speech through Smallest AI's Lightning models, running on Agno's own HTTP client so there is nothing extra to install. Another provider slotted into the toolkit surface rather than a change in what agents can do.

  2. 12d ago

    Poll the status of background metrics refreshes

    Adds a way to poll whether a background metrics refresh has finished. Plumbing for a job that already ran on its own schedule.

  3. 12d ago

    Run knowledge search on OpenSearch, with hybrid built in

    OpenSearch joins the supported vector stores with vector, keyword and hybrid search in sync and async form. The argument is consolidation: teams already running OpenSearch for logs can back a knowledge base with it instead of standing up a second store.

  4. 15d ago

    Ask your AgentOS how it's doing, in plain English

    ⚡ SPARK

    The platform becomes something an agent can interrogate about itself, which is a different posture than shipping a dashboard for humans to read.

  5. 15d ago

    Break down agent latency and errors in your traces

    Traces gain a breakdown of where latency and errors accumulate across an agent run. Useful once runs are long enough that a single duration number stops explaining anything.

  6. 15d ago

    Toggle thinking and feed files and video to Moonshot

    Moonshot gains a thinking toggle plus file and video inputs, bringing its integration closer to parity with the other model providers. Provider-surface catch-up, not a new capability for agents.