Yesterday's Trend Watch looked at the harness as a product. Today's sources look at what the harness carries between runs. An agent session is no longer one process from start to finish. It is paused and resumed, rewound, moved to a new machine, or opened by a different backend. Releases from October 29 to 31, 2026 show that every move needs its own test.
Moving between backends
Codex PR #58457, merged October 30, enables switching to a conversation owned by another backend in the TUI. Before, the session picker listed such conversations, but selecting one "was rejected." The change resumes the conversation through its owning backend and restores "the destination's settings, account state, model catalog, and workspace services." It blocks a switch during an active turn and keeps the current conversation if preparing the destination fails. Codex 0.166, covered here on October 30, had already started preserving approval settings across resume and fork.
Values that change type on resume
In ADK Python, a commit merged October 30 fixes workflows resumed from a JSON session store such as DatabaseSessionService. A finished node's bytes output "comes back as its base64 text." The fix decodes it when the node declares a bytes output, but "a producer with no return annotation still replays as base64 text." A live run and a resumed run of the same workflow could therefore see different types.
Choices lost after a cancel or a new machine
Gemini CLI PR #30059, merged October 30, fixes a selected model that "silently reverts" to a different model on later turns, including after cancelling a turn and running /rewind, while the footer and /model dialog still show the original choice. The fix keeps a temporary capacity fallback to the current turn. Claude Code 2.1.316 fixes a self-hosted runner losing "an MCP server's on/off switch" when a session moved to a new machine, a resume that did not restore a conversation's plan, and sub-agents switching to the main model when a background job woke them.
The same pattern at training scale
Sandman, Amazon's sandbox service for agentic reinforcement learning (arXiv 2610.39711), keeps files and processes alive across calls until the training framework releases the environment. Its authors list reclaiming abandoned sandboxes among their operational lessons. When sessions are long-lived and portable, cleanup and ownership become part of the design.
What remains uncertain. None of these release notes says how many users were affected. The fixes cover the paths their authors found, and other resume paths may lose state the same way.
Analysis: Treat every session transition as an interface with its own tests: pause and resume, rewind, fork, a move to another machine, and a switch to another backend. After each one, check that the model, permissions, tool switches, plan and typed outputs match what they were before. For the help-desk agent at Belgavi's Book Store, Geetha could pause a CASE-3150 refund run halfway, resume it on another worker and confirm that the agent still uses the approved model and the same refund amount type.
Sources and evidence
- Enable cross-backend session switching in the TUI (openai/codex #58457, October 30)
- fix: decode base64 bytes in rehydrated workflow node outputs (google/adk-python, October 30)
- fix(core,cli): preserve selected model across transient fallback, cancel, and rewind (google-gemini/gemini-cli #30059, October 30)
- Claude Code v2.1.316 release notes (GitHub, October 31)
- Sandman: Building and Operating Sandboxes for Agentic Reinforcement Learning on Shared Kubernetes (arXiv 2610.39711)
Releases from October 29 to 31 show agent sessions becoming portable state: Codex now resumes conversations across backends and restores their settings, ADK Python fixed bytes outputs replaying as base64 text after a JSON-store resume, Gemini CLI fixed a selected model silently reverting after cancel and rewind, and Claude Code fixed MCP switches and plans lost when sessions moved or resumed; every pause, resume, rewind and move needs its own test.