AI 程式代理
Prime Agent 0.9.2 Persists Trace Upload Progress with an On-Disk Cursor and Resumes Unfinished Work After Restart
The new release turns agent trace uploads into a persistent outbox. Processes no longer wait for network transfers when exiting, and can resume uploads from a byte cursor after restarting. It also begins recording semantic edges across sub-agents, but there is currently no corresponding delivery endpoint, so this should not be considered a complete distributed tracing capability.
Prime Intellect released Prime Agent 0.9.2 on September 5, addressing two of the areas most prone to failure in long-running, multi-agent workloads: trace data persistence, and agent-tree state and cost accounting. Previously, a process could wait for trace uploads when exiting. The new release uses an on-disk cursor outbox that stores each session’s upload intent and the position of content already sent. Scheduled uploads process only newly appended data. When rate-limited, they are rescheduled according to `Retry-After` instead of putting the process to sleep in place; after an unexpected termination, catch-up runs on the next startup.
This design decouples “agent task completion” from “successful telemetry delivery,” reducing the likelihood that network or tracing-service failures will hold up the daemon. The tradeoff is that the local outbox becomes a new form of persistent state: operators must monitor backlog size and disk usage, and clean up cursors after deleting sessions. Otherwise, asynchronous delivery merely postpones failures.
Version 0.9.2 also extracts a shared append-only event-log foundation that supports the RLM spawn ledger through single-call `O_APPEND` writes, optional fsync, bounded fail-closed replay, and a recovery mechanism that tolerates a truncated final line. The new `semantic-edges.jsonl` assigns stable request IDs to model requests and records parent-child sessions, sub-agent calls and returns, continuations, and compaction events, allowing tools to reconstruct multi-agent execution graphs. However, the release notes explicitly state that only derivation and outbox registration are currently implemented; there is not yet an endpoint that actually delivers this type of ledger.
Other updates make MCP server tools supplied by ACP clients natively callable, recursively aggregate sub-agent token usage and costs, and fix potentially exponential revisiting during cancellation of deeply nested agent trees. This is an execution-reliability update, with no new standalone capability benchmark; Prime Agent’s core still allows models to execute code through a persistent IPython kernel. Adopters should verify MCP permissions within an external sandbox, check whether trace content contains sensitive data, and test whether crash recovery can produce duplicate side effects.