What happened
Agno says it shipped 13 releases in October, while community members contributed 41 merged pull requests. The roundup’s central additions concern CompanyOS and the integration of existing coding agents into Agno’s AgentOS. CompanyOS is presented as a way to assign agents ongoing work that may otherwise be neglected, such as keeping documentation current or triaging GitHub issues. A team configures AgentOS with a database and defines Products and the Sources their agents should consult. Routines read those Sources and propose Projects and Tasks. People approve the proposed work before agents carry it out, and completed Tasks return for review.
Agno also introduced CodexAgent, which wraps the official openai-codex SDK, alongside its ClaudeAgent. The coding agents retain their own execution loops: they run shell commands, edit files, search the web, and call MCP tools. Agno supplies surrounding session and AgentOS API functionality. Background runs return a run_id that clients can poll, with a cancel endpoint to interrupt a turn. Callers can provide a Pydantic model or JSON Schema as output_schema, after which run.content is returned parsed. Claude Code and Codex can be assigned different roles in a workflow—for example, one writes a test and the other reviews it. Native permission requests become paused runs requiring confirmation in the AgentOS chat.
For persistence, ClaudeAgent and CodexAgent can each receive a sandbox. Agno runs the coding agent as a process in its own computer environment and checkpoints workspace and conversation files after every turn. If a container dies, the stated recovery point is the last completed turn. Docker and Amazon Bedrock AgentCore are named as providers; checkpoints can be stored in S3 or an existing Postgres database, and another replica can resume a session. A repository can be attached with a branch, whose commit is pinned on the session’s first turn.
Other operational changes include context compaction, authorization, and filesystem tools. With compaction enabled, Agno waits for the provider to reject an oversized request, summarizes older turns with the agent’s model, and retries with the summary plus as many as the five most recent runs. Teams may instead set a compaction threshold or call compact() directly. AgentOS Authorization verifies a caller’s JWT and checks route scopes against roles in its database. The durable filesystem gives the agent seven file tools backed by an agno_fs table; user_isolation can restrict each user to their own files. The roundup also reports that a worker wake-up change reduced submit-to-RUNNING time in a ten-run stub-model test from 91–763 milliseconds to 17–53 milliseconds.
The roundup flags migration and behavior changes: DbFileSystem users must run a migration while the application is stopped; MCPConfig(tools=[...]) now publishes only listed tools; and CodexAgent with a database but no approval settings routes each review to a person, potentially pausing the run. Agno advises reading release notes before upgrading.
What the archive adds
The archive has covered runtime controls for coding agents in [Adversa AI’s coding-agent security platform] (https://newruntime.com/posts/adversa-ai-launches-coding-agent-security-platform-for-282ec1/), but that story focused on monitoring and blocking actions, not Agno’s session and recovery features. This is a new fact in the broader coding-agent operations story, not a restatement of that product announcement.
Trend signal
This supports Harness architecture outlives model choice: Agno’s value proposition here is session orchestration, persistence, approvals, and recovery around agents that retain their own execution loop. The roundup does not establish that teams prefer this architecture or that it performs better in production.
Why now
The roundup is dated November 4, 2026 and covers releases shipped during October. It frames the work around two operational gaps: agents need a way to identify and execute approved ongoing tasks, and coding-agent sessions need to survive host loss. The source gives no external market trigger for the timing.
Why it matters
A coding agent’s model is only one part of a production workflow. Teams also need to decide where files and conversations live, what happens when a process or container fails, how a person approves sensitive actions, and how another worker resumes work. Agno’s approach exposes those concerns as framework capabilities rather than leaving all of them to application code. That can help teams evaluate whether they want a shared layer around Claude Code and Codex while preserving each tool’s own command and editing loop.
The details also make the trade-offs visible. Checkpoints are described per completed turn, not continuously, so work inside a failed turn may not be recoverable. Approval defaults can pause runs, which is safer than silently proceeding but affects unattended workflows. Persistent agent files and role-based route scopes introduce data-isolation and authorization design decisions that teams must test in their own deployment. The reported queue latency improvement is a small stub-model test, not evidence of end-to-end production speed.
What happens next
- Teams adopting the coding-agent integrations will need to document approval behavior and recovery points before relying on unattended runs. (3 months, 70%). Watch for: deployment guidance that specifies which state survives a failed turn and how paused approvals are resumed.
Limits and uncertainty
This is a vendor-authored roundup, and the claims are not independently evaluated in the supplied material. It does not provide benchmarks for coding quality, recovery reliability, security, or production-scale throughput. The latency result covers ten runs with a stub model and measures submit-to-RUNNING time, not task completion. Agno names supported providers and storage options but gives no pricing, deployment limits, or compatibility matrix. The roundup also does not quantify adoption or show whether CompanyOS routines reliably identify useful work. Its descriptions establish what Agno says the framework can do, not whether these features suit a particular team’s governance or reliability requirements.
Skill to build now
For one non-production coding-agent workflow, test the failure boundary: run a task in a sandbox, complete a turn, force the container to stop, and verify what files and conversation state another replica can resume. Separately, trigger a permission request and confirm whether the run pauses as expected. This directly checks the two operational claims that matter most before a team depends on the integration.
Reader takeaway
Agno is adding an operational layer around Claude Code and Codex rather than replacing their execution loops. The most consequential details are turn-level checkpoints, explicit human approvals, and persistent state; teams should verify those boundaries in their own environment before treating a coding-agent session as durable.