Claude Code Routines put agents on a schedule

Anthropic introduced Routines in Claude Code as a research preview. A routine combines a prompt, repository, and connectors, then runs on hosted infrastructure according to a schedule, an API call, or an event.

That takes a coding agent beyond the interactive terminal session. Since Anthropic hosts the runtime in the cloud, a routine can keep going without a developer’s laptop staying open. Issue triage and routine repository maintenance are obvious uses.

It also makes failure handling harder to ignore. Scheduled tasks need bounded permissions, useful logs, and some policy for repeated failures. Anything that may be retried should be idempotent. Super-charged and super-capable cloud cron jobs!

A practical inventory of coding-agent parts

Sebastian Raschka published a walkthrough of the components of a coding agent, including repository context, tool use, memory, and delegation.

The walkthrough separates the model and terminal loop from the surrounding choices: which files get selected, which commands are available, what persists, and when another agent gets involved. That decomposition is useful to understand the components of a coding agent system and also a map of where to look when the coding agent is struggling or going off the rails.

Advisor Mode: Opus advised; Sonnet or Haiku executed

Anthropic also describes an advisor strategy that pairs Opus as an advisor with Sonnet or Haiku as the executor.

In this design, the stronger model handles planning or review and a cheaper model handles execution. Anthropic reported near-Opus intelligence at lower cost. I am curious how often the executor notices that a good plan has stopped fitting the work in front of it.

Three to five agents, then attention gets scarce

A post about OpenAI’s internal experience said engineers could supervise roughly three to five coding agents before productivity dropped. It also said Symphony was open-sourced to reduce that supervision bottleneck.

The number will vary with the task and interface. Every agent can produce questions, diffs, alerts, and failed runs, all competing for the same person’s attention. Orchestration becomes a human-attention and human-factors problem. Suspect the tools themselves have to evolve to help users (human orchestrators) manage this better.

Comparing what agents keep and discard

This writeup compares file reading, compaction, and sub-agent strategies across Pi, OpenClaw, Claude Code, and Letta.

Context management is usually hidden behind product behavior. Two agents using similar models can feel quite different depending on when they compact, what they retain, and how they delegate subproblems.

There probably isn’t one strategy that fits coding, research, and personal-assistant workloads; each has a different memory shape. The comparison focuses on file reading, compaction, and delegation rather than context-window size alone.

A repository map for structural context

repowise indexes a codebase into structural information meant to show dependencies, coupling, and ownership.

Most coding agents search by filename and text because those are the interfaces available to them. A repository graph can expose relationships that are difficult to infer from a handful of retrieved files. Its usefulness will depend on language support and the indexing model, though even an incomplete map may stop an agent from changing a central file without seeing how widely it is used.

HALO uses traces to propose runtime changes

The Hierarchal Agent Loop Optimizer, or HALO, uses a language model to analyze execution traces and repeatedly suggest changes to the agent.

The traces become input to a process that proposes changes to instructions or runtime behavior. Need to see a lot more examples of the changes HALO suggests to figure out how useful it actually is.