Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

the agent log grew joints

LLM 0.32 separates reasoning, tool activity, approval pauses, and stored history into inspectable parts instead of one opaque chat log.

Published Updated Read 2 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-08-05

an agent that uses tools leaves more than an answer behind: reasoning, tool calls, approvals, and a history somebody may need to reopen. LLM 0.32 puts those pieces on separate surfaces instead of stuffing the whole mess into one chat-shaped log.

1. the agent log grew joints

sources:

what happened: LLM 0.32 adds streamed event types for reasoning, text, tool calls, and attachments. Reasoning traces now go to standard error, leaving ordinary output clean for a pipe; the release also adds provider-side tools, including code execution and web search, plus tool chains that can pause for human approval and resume from stored message history. Underneath, its SQLite logs now store messages by content address, Git-style, so repeated turns do not keep writing the same growing transcript as duplicate JSON.

the collision: this is a small CLI release with a more adult idea inside it: agent work needs separable evidence. A clean answer is useful, but it is not the same thing as the reasoning stream, the tool that touched a system, or the moment a person chose to let the loop continue. Keeping those objects distinct makes an assistant easier to inspect, replay, and refuse without asking the owner to excavate a single very long scroll.

question left open: when a tool loop stops for approval, what is the smallest record a person needs before “continue” counts as an informed decision?

left on the table

  • Agent Reach has real operational detail around local cookies, browser sessions, fallback backends, and a diagnostics command, but it is still a fresh integration catalogue rather than evidence that the access boundary itself changed.
  • Rudder links prompt history to branch-level tests and local SQLite storage; a good object, but it would reopen last week’s test-rig and review lane for a young repository claim.
  • Guardian Angels is a serious proposal for highly personalised models, but a long argument about a future system is not the same as a newly verifiable operating surface.
agent-operations audit-logs human-approval personal-ai