Skip to content

blog: When Your Coding Agent Goes Down, Don't Start Over - #4087

Draft
benfrank241 wants to merge 1 commit into
mainfrom
blog/switch-coding-agents
Draft

blog: When Your Coding Agent Goes Down, Don't Start Over#4087
benfrank241 wants to merge 1 commit into
mainfrom
blog/switch-coding-agents

Conversation

@benfrank241

Copy link
Copy Markdown
Member

Makes the case for repo-owned rather than agent-owned memory, using provider outages as the entry point.

The argument

An outage doesn't cost you the agent, which is replaceable in seconds. It costs you the context you built up inside it, which isn't, because it only ever existed in one place.

Each agent keeps its own session history in its own format, and its own project-context convention (CLAUDE.md, QWEN.md, AGENTS.md). None reads the others'. So the agent you switch to starts where the first one started on day one.

Since all 16 supported agents share one bank per repository by default (coding-agent::{gitProject}), switching is just switching. What Claude Code learned this morning is what Codex reads this afternoon.

On in-flight sessions

The post does not claim seamless failover, because that isn't what happens:

  • Plugin harnesses (opencode, opencode 2, Kilo) upsert the transcript every turn, plus an idle flush.
  • Hook harnesses (Claude Code, Codex, Qwen Code, most others) write on Stop.
  • The background history import picks up transcripts that never reached a Stop hook.

So a clean exit loses nothing, and a hard crash costs the current turn rather than the session. That's stated plainly in its own section — it's a strong enough position without overclaiming, and it's the first thing a skeptical reader would test.

Scope

Closes by widening past outages to the cases that come up more often: rate limits, choosing the right agent per task, evaluating a new agent without losing a day to onboarding, and teams where two engineers use different agents against one repo.

Verification

  • 16 harnesses and the shared-bank default confirmed against the v0.5.1 package README.
  • Write-back behavior per harness type confirmed against the retainSessions documentation.
  • All three cross-links verified live.

Content score

Composite 88.5/100 (humanity 100, specificity 94, readability 94).

🤖 Generated with Claude Code

https://claude.ai/code/session_01AE4NCnCKXUzNDF3N9FaQja

Makes the case for repo-owned rather than agent-owned memory, using
provider outages as the entry point.

The argument: an outage does not cost you the agent, which is replaceable
in seconds. It costs you the context you built up inside it, which is not,
because it only ever existed in one place. Agents each keep their own
session store and their own project file convention, so the one you switch
to starts from day one.

Since all 16 supported agents share one bank per repository by default,
switching is just switching. What Claude Code learned in the morning is
what Codex reads in the afternoon.

Includes an honest section on in-flight sessions rather than claiming
seamless failover: plugin harnesses upsert every turn, hook harnesses
write on Stop, and the background history import picks up transcripts that
never reached a Stop hook. A hard crash costs the current turn, not the
session.

Closes by widening past outages to rate limits, per-task tool choice,
evaluating a new agent, and teams on mixed agents.

Cover is k03: eight agent slots, one lit, all converging on one memory.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AE4NCnCKXUzNDF3N9FaQja
@benfrank241
benfrank241 marked this pull request as draft September 3, 2026 20:59
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant