Bilingual

Show HN: Mirrors – test AI agent changes by replaying production traces
19 days ago
- Mirrors converts production traces into an isolated copy of an agent's environment for safe testing.
- It rebuilds schemas, discovers tools, and provides a runnable copy with seeded databases and scored tool matches.
- Enables replaying agents consistently to measure accuracy, catch regressions, and ensure changes work before deployment.
- Offers deterministic, byte-identical worlds for reliable failure reproduction without touching production systems.
- Supports on-demand launch, scaling to zero, per-minute metering, and integration via a versioned API and workspace keys.
- Provides free building with deterministic seeding, with Custom plans for unlimited sandboxes, API access, and SSO.
Cloud Agent Lessons
13 days ago
- Cloud agents evolved from simple extensions of local agents to independent VMs with their own environments, enabling parallel, unattended, and long-running tasks.
- A full development environment is critical for cloud agent output quality, requiring infrastructure like user tools, VM hibernation, and pipelines for checkpointing and restoring images.
- Cloud agents face reliability challenges due to disruptions like inference outages; using Temporal for durable execution improved reliability and now handles millions of actions daily.
- Decoupling agent loops, machine state, and conversation state allows flexible deployment across different pods, with efficient storage for streaming updates and handling retries.
- The harness for cloud agents shifted from tightly controlling behavior to giving agents more autonomy via tools, adapting as models got smarter, while still providing scaffolding for complex tasks.
- Future focus includes self-healing environments where agents can report and fix issues like missing secrets or network blocks, aiming to reduce manual intervention.
Zenoh Reliability, Scalability and Congestion Control
10 hours ago
- Reliable many-to-many communication over wide area networks is challenging, requiring trade-offs between reliability, scalability, and resource consumption.
- zenoh's protocol has two layers: a session protocol for 1-to-1 communication with best-effort and reliable channels, and a routing protocol to propagate data from producers to consumers.
- Hop-to-hop reliability is scalable and reliable under stable infrastructure but may lose data during topology changes.
- Stronger reliability options include end-to-end (no loss during topology changes but less scalable) and first-router-to-last-router (reduces pressure on endpoints).
- Flow control separates control over resending (receiver), memory usage (sender/infrastructure), and message dropping (sender congestion control).
- Configuration examples show reliability set on the subscriber side and congestion control per sample using the Python API.
- zenoh allows balancing application needs with system constraints, protecting nodes from asymmetries and maintaining strong guarantees in dynamic networks.

Hasty Briefsbeta

#reliability

Show HN: Mirrors – test AI agent changes by replaying production traces

Cloud Agent Lessons

Zenoh Reliability, Scalability and Congestion Control