Dev.to AI ๐Ÿค– Ai ๐Ÿ‘ 0 ๐Ÿ“– 7 min read

Building More Reliable AI Systems: RagLeap Core v0.13.6 and Observability 0.8.6

AI systems are becoming increasingly capable, but capability alone does not make a system reliable. An AI agent may have access to multiple models, tools, memory systems and specialized sub-agents. Yet a simple orchestr

AI systems are becoming increasingly capable, but capability alone does not make a system reliable.

An AI agent may have access to multiple models, tools, memory systems and specialized sub-agents. Yet a simple orchestration failure can prevent it from completing a task. A slow model provider can delay an entire workflow. Meanwhile, the infrastructure supporting these systems must remain observable, upgradeable and dependable.

At RagLeap, we believe reliable AI infrastructure requires more than adding new capabilities. It requires continuously testing how those capabilities behave under real operating conditions.

Two recent releases address different parts of this challenge:

  • RagLeap Core v0.13.6: Improvements to sub-agent task completion and provider timeout handling.
  • ragleap-observability 0.8.6: Updated Prometheus, Loki and AlertManager images, with Kubernetes compatibility, alert testing and security scanning.

This article explains what changed, why it matters, what was tested and what remains unverified.

1. RagLeap Core v0.13.6: Improving AI Agent Orchestration

GitHub release: https://github.com/ragleap/ragleap-core/releases/tag/v0.13.6

Version 0.13.6 introduced two fixes discovered during the first real-model test of sub-agents on a live server.

These were not simply theoretical edge cases. They exposed practical orchestration problems that became visible when actual model providers were involved.

Problem 1: Sub-agents that completed no tool actions returned no useful answer

Consider a multi-agent workflow:

  1. A parent agent delegates a task to a specialized child agent.
  2. The child evaluates the task.
  3. The child determines that no external tool is necessary.
  4. The child returns its execution result to the parent.

Previously, when the child required no tool, it returned:

0 step(s): no actions

The parent received no useful task answer and could delegate the same task again.

This created an unnecessary orchestration loop.

What changed in v0.13.6?

PR #1473 introduced a short model call that allows the child agent to formulate an answer even when no tool execution is required.

The resulting answer is returned to the parent agent.

This improves an important distinction in multi-agent systems:

A task can be completed through reasoning alone. The absence of tool execution should not automatically mean the absence of a useful result.

PR: https://github.com/ragleap/ragleap-core/pull/1473

Problem 2: Slow model providers could stall agent planning

Multi-provider support gives an AI system flexibility, but provider availability and response times are not always predictable.

During testing, a Gemini overload response (503 high demand) caused one delegation to take almost 12 minutes.

The next provider answered in approximately five seconds.

That difference demonstrated how one slow provider could delay an otherwise successful workflow.

What changed in v0.13.6?

PR #1474 introduced:

ACTION_PROVIDER_TIMEOUT

Default: 45 seconds.

The setting controls the time limit for each provider's planning or summary call. When the limit is reached, the workflow moves to the next provider.

Setting the value to 0 disables the timeout.

The objective is straightforward: a slow provider should not indefinitely hold up agent planning when another provider may be available.

PR: https://github.com/ragleap/ragleap-core/pull/1474

An important limitation

An abandoned provider call may continue running in the background until the provider library returns. Its result is ignored, but the underlying call is not necessarily cancelled immediately.

The release also had not yet been validated against a real provider outage or a real-model sub-agent run after these fixes. That follow-up testing was planned after the release.

Safe upgrade behaviour

Version 0.13.6 includes:

  • No database schema change.
  • No mandatory new configuration.
  • File writing remains disabled.
  • File deletion remains disabled.
  • Sub-agents remain disabled by default.

Upgrading does not automatically enable these capabilities.

This matters because improving orchestration should not silently change an operator's security or autonomy settings.

2. ragleap-observability 0.8.6: Strengthening Monitoring Infrastructure

GitHub release: https://github.com/ragleap/ragleap-core/releases/tag/ragleap-observability-v0.8.6

Monitoring is essential for operating AI infrastructure. It is not enough to deploy Prometheus, Loki and AlertManager. Their images, dashboards, alert rules, data sources and upgrade behaviour must also be tested.

Version 0.8.6 updates three core monitoring components.

AlertManager v0.35.0

Previous version: v0.34.1

The updated image passed its Trivy image scan gate with:

  • 1 HIGH finding.
  • 0 CRITICAL findings.
  • Down from 7 HIGH findings in the previous comparison.

The updated chart was tested through installation, upgrade and uninstall on:

  • Kubernetes v1.34.11
  • Kubernetes v1.36.4
  • Kubernetes v1.37.0

The end-to-end alert test also passed all checks.

Loki 3.7.10

Previous version: 3.6.17

The updated Loki image passed its Trivy image scan gate.

Validation included:

  • Chart installation.
  • Chart upgrade.
  • Chart uninstall.
  • Kubernetes v1.34.11, v1.36.4 and v1.37.0.
  • Loki datasource health checks after installation and upgrade.
  • End-to-end alert testing.

All reported checks passed.

Prometheus v3.16.0

Previous version: v3.13.4

The updated Prometheus image passed its Trivy image scan gate with:

  • 1 HIGH finding.
  • 0 CRITICAL findings.

The chart was installed, upgraded and uninstalled across the same three Kubernetes versions.

The end-to-end alert test verified:

  • Four alert rules loaded and healthy.
  • A fast-burn alert fired during an outage and subsequently cleared.
  • RagleapTargetAbsent fired when a monitored target disappeared.

The tests/promtool rule unit tests were also updated to run against Prometheus v3.16.0.

This is important because monitoring rules should be tested against the version operators are actually deploying, rather than relying indefinitely on an older test image.

3. Testing the Combined Monitoring Stack

Testing each component individually is useful, but production infrastructure runs components together.

For that reason, the release work also tested all three updated images in one chart using the throwaway tag:

test-rel-observability-0.8.6

The combined validation covered:

  • Chart installation.
  • Chart upgrade.
  • Chart uninstall.
  • End-to-end alert testing.
  • Trivy scanning of the three updated images.

The installation, upgrade, uninstall and end-to-end alert tests were successful.

An important security-scanning distinction

The combined Trivy run passed for all three updated images.

However, the only failing job was for neo4j:5-community, which is not deployed by this chart.

The failure was associated with a CRITICAL finding in the Debian linux-libc-dev package:

CVE-2026-99306

Advisory: https://github.com/advisories/GHSA-2p6h-wr8g-2xqx

The same issue also fails on the main branch.

It is important to report this accurately: the three updated monitoring images passed their gates, but the combined scan was not entirely clean.

Security scanning should communicate both successful checks and outstanding findings.

4. Packaging and Distribution

The observability release also includes packaging verification.

The published PyPI wheel for version 0.8.6 was verified to contain 17 files under the helm/ directory.

The project source URL was also verified.

Install the package:

pip install ragleap-observability==0.8.6

Package: https://pypi.org/project/ragleap-observability/

Source code: https://github.com/ragleap/ragleap-core/tree/main/packages/ragleap-observability

5. What We Have Not Claimed

Engineering transparency means identifying what has not been tested.

For ragleap-observability 0.8.6:

  • Alloy remains at v1.20.1. Although v1.22.0 scanned cleanly, no test currently enables Alloy, so it was not upgraded.
  • The manual live SLO test (tests/live/slo-live-test.sh) was not rerun. It still pins Prometheus v3.13.4.
  • Upgrades of long-running clusters containing existing Prometheus, Loki or AlertManager data were not tested.
  • Actual Slack or email notification delivery remains unconfirmed.

For RagLeap Core v0.13.6:

  • A real provider outage was not yet tested after the timeout fix.
  • A real-model sub-agent run after both fixes was still pending.
  • A timed-out provider call may continue in the background until its library returns.

These limitations are part of the release record, not details to hide.

6. Two Different Layers, One Engineering Goal

These releases address separate engineering problems.

Engineering area Release Improvement
Agent orchestration Core v0.13.6 Child agents return useful answers without requiring tools
Provider resilience Core v0.13.6 Slow planning and summary calls can fall back to another provider
Monitoring security Observability 0.8.6 Updated images with Trivy validation
Kubernetes compatibility Observability 0.8.6 Installation, upgrade and uninstall testing across three Kubernetes versions
Alert reliability Observability 0.8.6 End-to-end alert and Prometheus rule validation

A dependable AI platform needs both intelligent orchestration and dependable infrastructure.

An agent that can reason, delegate and use tools still depends on reliable model providers, monitoring, alerting and operational controls.

Likewise, a well-monitored infrastructure stack does not automatically make an AI workflow reliable. Both layers must evolve together.

7. The Bigger RagLeap Direction

RagLeap is an open-source, self-hosted AI platform focused on building AI employees and the infrastructure around them.

Our broader direction goes beyond individual LLM applications or retrieval pipelines.

We are working towards an ecosystem in which AI employees can use specialized tools, coordinate tasks, maintain context and eventually interact across software, voice, browsers, computers and physical environments.

That vision requires strong foundations:

  • Reliable agent orchestration.
  • Multi-provider resilience.
  • Memory and contextual retrieval.
  • Security and permission boundaries.
  • Observable infrastructure.
  • Controlled execution and human oversight.

Our Physical AI work is currently at the simulation stage. We are approaching real-world hardware interaction carefully, with safety and authorization as foundational requirements.

We do not consider a feature complete simply because code has been written. Testing, validation and operational evidence matter.

8. Open Source Is a Collaborative Engineering Process

RagLeap is built in the open.

We welcome developers, AI engineers, infrastructure engineers, security researchers, robotics enthusiasts and contributors interested in solving real engineering problems.

There are meaningful challenges across agent orchestration, provider integration, memory, security, observability, Kubernetes, voice and Physical AI.

You do not need to work on every part of the platform. Focused contributions, reproducible bug reports, documentation improvements, tests and thoughtful technical reviews all matter.

If you are interested in building reliable AI systems rather than just experimenting with model prompts, we would be glad to have you explore the project.

Explore RagLeap:

Final thought

The future of AI will not be determined by model capability alone.

It will also depend on how reliably AI systems reason, delegate, recover from failures, interact with infrastructure and operate within safe boundaries.

That is the engineering challenge we are working on at RagLeap โ€” one tested improvement and one release at a time.

๐Ÿ“ฐ Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ€” full credit and traffic to the original publisher.