Runeval
2 subscribers
12 links
Practical AI research and guides to help you choose tools, understand tradeoffs, and build reliable systems. Explore the full articles at https://runeval.dev
Download Telegram
Channel created
Welcome to Runeval 👋

Practical AI research for developers choosing tools and building reliable systems.

Expect useful guides, clear tradeoffs, and implementation pitfalls drawn from documentation and research.

Start with our practical guide to agent memory, checkpoints, and durable workflows. Then explore the tool reports that match your work.

Browse the library: https://runeval.dev/?utm_source=telegram&utm_medium=social&utm_campaign=channel_welcome&utm_content=pinned

By alokrm
Runeval pinned «Welcome to Runeval 👋 Practical AI research for developers choosing tools and building reliable systems. Expect useful guides, clear tradeoffs, and implementation pitfalls drawn from documentation and research. Start with our practical guide to agent memory…»
PRACTICAL GUIDE · Agent reliability

Agent memory, checkpoints, and durable workflows: choosing what to persist

A checkpoint can preserve progress without proving that an external action happened only once. If a worker loses a refund receipt, recovery must resolve the provider’s outcome before issuing another action.

For developers designing agents that pause, resume, or call external services. Includes a worked refund scenario, two diagrams, comparison tables, and recovery checks.

15 min read · By alokrm

https://runeval.dev/guides/agent-memory-checkpoints-durable-workflows?utm_source=telegram&utm_medium=social&utm_campaign=agent-memory-checkpoints-durable-workflows&utm_content=initial
AGENT STATE · LangGraph

When does an agent need persistent state?

Conversation history, execution progress, and long-term memory have different jobs. Treating them as one growing transcript makes recovery and retrieval harder to reason about.

For developers whose agents span interruptions, approvals, or multiple sessions. This report explains the adoption threshold and the recovery questions to answer first.

3 min read · By alokrm

https://runeval.dev/tools/langchain-ai-langgraph?utm_source=telegram&utm_medium=social&utm_campaign=langchain-ai-langgraph&utm_content=initial
DURABLE JOBS · Vercel Workflow

When should a TypeScript job become a workflow?

Start with the failure it must survive. Saved progress and retries help a job continue after interruptions, while external writes still need their own duplicate-request protections.

For TypeScript developers building jobs with long waits or several stages. The report covers workflow and step boundaries, backend ownership, and why waiting without compute is not a complete cost model.

3 min read · By alokrm

https://runeval.dev/tools/vercel-workflow?utm_source=telegram&utm_medium=social&utm_campaign=vercel-workflow&utm_content=initial
EVALUATION · Promptfoo

Build checks that catch real regressions

A higher aggregate score can hide a serious failure. Define the release decision first, then keep output-shape checks, semantic criteria, and known regression cases visible as separate requirements.

For developers comparing prompts or models and adding evaluation to CI. The report explains useful comparison boundaries, model-judge limitations, and how regression gates differ from red teaming.

3 min read · By alokrm

https://runeval.dev/tools/promptfoo-promptfoo?utm_source=telegram&utm_medium=social&utm_campaign=promptfoo-promptfoo&utm_content=initial
AI agent observability

MLflow for AI Agents: Tracing, Evaluation Metrics, and Gateway Architecture

MLflow treats an agent run as a hierarchy of OpenTelemetry spans, so teams can inspect tool calls, retrieval context, latency, exceptions, and final output in one trace, then evaluate those traces for correctness, relevance, safety, and other failure modes. The trade-off is operational: high-volume traces increase storage and serialization load, while LLM judges are usually better run asynchronously on sampled batches than on every production span.

Useful for platform and ML engineering teams deciding whether they need one observability and evaluation layer across agent systems and traditional ML workloads.

5 min read · By alokrm

Read the article →
Distributed vector search infrastructure

Milvus Vector Database: How Compute-Storage Separation Scales ANN Search

Milvus separates query execution from ingestion, so read-heavy traffic can scale Query Nodes without scaling the write path, while bulk ingestion can expand Data Nodes independently. That flexibility comes with Kubernetes and multi-service operational overhead in a distributed deployment; Milvus Lite offers a file-backed Python path for local prototypes.

Useful for platform and ML engineers evaluating vector databases for high-scale RAG or independently scalable read and write workloads.

4 min read · By alokrm

Read the article →
Durable TypeScript AI agents

How Trigger.dev Sustains Long-Running AI Agents via Checkpointing and Zero-Idle Waits

A long-running agent does not need to keep a worker occupied while it waits. Trigger.dev checkpoints the run at a delay, webhook callback, or human approval, frees the compute, then restores the saved execution state when work resumes.

Useful for TypeScript developers and platform engineers designing approval-gated agents, durable background jobs, or workflows that wait on external events.

4 min read · By alokrm

Read the article →
Parallel AI coding agent orchestration

Orchestrating Parallel AI Coding Agents with amux's Rust Control Plane

A shared task board can avoid duplicate agent work by claiming each task with an atomic compare-and-swap operation. amux pairs that ownership rule with separate done and verified gates, so implementation and peer or CI validation remain distinct instead of being collapsed into one status.

Useful for engineering teams evaluating local-first orchestration for parallel AI coding agents.

5 min read · By alokrm

Read the article →
Unified graph and vector storage

Can HelixDB Unify Vector Search and Graph Traversals on Object Storage?

HelixDB Cloud pairs object-storage-backed persistence with one writer and auto-scaling readers. That lets teams expand read capacity for retrieval workloads, but write-heavy ingestion remains bounded by the single-writer topology—a key tradeoff to test before replacing separate graph, vector, and application databases.

Useful for platform and AI engineers evaluating a unified data layer for RAG, agent memory, or knowledge-graph applications.

5 min read · By alokrm

Read the article →
AI application performance

Why AI responses are slow: tracing retrieval, tools and model generation

Measure two waits separately: user submit to the first visible answer, and user submit to a complete, usable result. A fast first protocol chunk may contain only metadata, while a prompt first answer can still be followed by a slow completion. Trace one request across application queues, retrieval, every tool attempt, every model call and display so you optimize the measured interval that actually delays the user milestone—and verify the end-to-end result improves without more errors or timeouts.

For engineers and technical leads diagnosing latency in retrieval-augmented, tool-using or streaming AI applications.

13 min read · By alokrm

Read the practical guide →
Open-source LLM observability

Opik: Technical Architecture for Open-Source LLM Observability

Opik can keep one evaluation workflow across development and production: teams can compare prompts or models against datasets, enforce PyTest checks in CI, and run asynchronous online evaluation rules on production traces. The self-hosted Apache-2.0 stack adds infrastructure responsibility—Docker Compose locally or Kubernetes and Helm at larger scale—while its managed cloud option trades that control for lower operational overhead.

Useful for ML and platform engineers choosing between self-hosted and managed tracing, evaluation, and monitoring for LLM applications and agents.

3 min read · By alokrm

Read the article →