Hermes Agent v0.19 Survives Crashes and Resumes Interrupted Tasks Mid-Turn
Hermes Agent v0.19 shipped a major reliability update late last week, merging 99 pull requests in a single day. The headline feature is crash resilience: agents can now survive app crashes, backend failures, and machine reboots without losing work.
What crash resilience means
Previous versions of Hermes could recover a transcript after a crash, but they could not resume from the exact point of interruption. If your agent was on turn 12 of a 20-turn task and the process died, you started over.
v0.19 changes the recovery model. When a long-running task is interrupted:
-
Exact turn recovery. Hermes resumes the interrupted turn, not just the conversation history. It knows precisely which tool call or reasoning step was in progress.
-
Completion inspection. Before continuing, it checks what was already done -- files written, API calls made, state committed -- and avoids duplicating work.
-
Crash-loop breaker. If the same error repeats at the same turn, Hermes detects the loop and stops retrying. This prevents the agent from burning tokens on an unrecoverable failure path.
Hermes now automatically resumes the exact interrupted turn (not just recovering a transcript), inspects what was already completed, and picks up and finishes the job without duplicating work.
Durable response delivery
A related improvement addresses a common failure mode in production: the gateway crashes between generating a response and delivering it to the user's channel.
Hermes now writes finished responses to persistent storage before attempting delivery. If the gateway goes down before the response reaches Telegram, Discord, or Slack, the message is re-delivered on the next boot. The user sees the response as if nothing happened.
Performance improvements
The v0.19 release also delivered significant speed gains. The following benchmarks were reported by the community:
| Metric | Before | After | Improvement |
|---|---|---|---|
| Cold start | ~4.3s | ~0.9s | 79% faster |
| Reply latency | Baseline | ~80% faster | ~80% |
| Skill discovery | Baseline | 5x | 5x |
| Worktree startup | Baseline | Up to 8.7x | Up to 8.7x |
These improvements make short tasks practical in a way they were not before. A quick file check or status query that took several seconds of agent overhead now completes in under a second.
Stability work under the hood
Beyond the user-facing features, the release included substantial internal stability fixes:
- SQLite runtime hardening and state.db recovery paths
- Checkpoint and orphan session management
- Memory consistency and compression improvements
- Session lineage tracking across crashes
- Tool-call recovery for interrupted tool executions
- UTF-8 corruption fixes
- Prevention of premature auto-deletion of active sessions
- Lease conflict resolution
Backup and migration
v0.19 also ships a backup system that works while the agent is running:
hermes backup # Full backup with timestamp
hermes backup --quick # Faster, skips some compression
hermes import <backup.zip> # Restore on a new machine
Backups include config, memories, sessions, skills, cron jobs, and databases. The tool uses SQLite's backup API, which is safe to run on live databases. This makes it practical to migrate an agent between machines or duplicate a configuration for testing.
Why this matters for production
Agent reliability has been a gap between "demo" and "production" deployments. A 30-minute autonomous run that fails at minute 28 because of a transient network error or an OOM kill means starting over from scratch. The v0.19 improvements change the calculus:
- Autonomous runs of 30-40 minutes are now practical on commodity hardware
- Machine reboots no longer destroy in-progress work
- Gateway crashes are invisible to end users
- Repeated failure patterns self-terminate instead of burning tokens
The combination of crash resilience, durable delivery, and faster cold starts narrows the gap between running an agent locally as a curiosity and deploying it as a production service. For teams already running Hermes on VPS instances, upgrading to v0.19 is a straightforward reliability gain.
[^1]: Tony Simons [@tonysimons_]. Thread on Hermes v0.19 crash resilience and turn recovery. [^2]: Julian Goldie [@JulianGoldieSEO]. Durable response delivery and v0.19 reliability improvements. [^3]: Julian Goldie [@JulianGoldieSEO]. v0.19 performance benchmarks and 99-PR merge day. [^4]: Luke [@iamlukethedev]. Stability edge case fixes: orphans, UTF-8, lease conflicts. [^5]: witcheer [@witcheer]. Hermes backup and import commands.