🔥 Story of the Day
Kore counts down to Artemis, its moonshot for governable AI agents (Kore Agent Platform) — The New Stack
Kore released Artemis, the latest edition of its Kore Agent Platform, which aims to structure enterprise agent development beyond basic prompt-chaining. The platform provides a visual and code-based environment for building, governing, and optimizing multi-agent AI systems. The core innovation is the declarative blueprint language, the Agent Blueprint Language (ABL).
ABL standardizes the definition and governance of AI workflows by treating them as compiled, declarative artifacts rather than sequences of imperative calls. It hardcodes support for complex orchestration patterns, including supervisor oversight, delegation, and agent-to-agent federation. This moves the paradigm away from the purely imperative nature of existing prompt-chain frameworks.
This maturation of the definition layer is critical for production ML infrastructure. By providing a declarative contract for how agents interact, it imposes the necessary governance and portability layer required to move multi-agent systems beyond brittle, ad-hoc scripting into reliable, deployable software components.
⚡ Quick Hits
OpenTelemetry graduates into the AI infrastructure era (OpenTelemetry Hits General Availability) — The New Stack
OpenTelemetry's elevation to CNCF graduation solidifies its position as the stable, standardized observability backbone. This maturity is significant because it guarantees a neutral, vendor-agnostic foundation for collecting traces, metrics, and logs.
For ML workloads, especially those running complex, self-hosted services like LLMs on Kubernetes, this stability is paramount. It confirms the tooling layer necessary to monitor complex state transitions and distributed inference patterns is mature and battle-tested across the industry.
How NetEase Games achieved 30-second LLM cold starts on Kubernetes (CNCF Blog) — CNCF Blog
Operating large language models at scale exposes that the primary performance constraint is data loading, not scheduling capacity. Traditional autoscaling mechanisms fail if model weights cannot be provisioned rapidly to the compute nodes.
NetEase mitigated this by implementing a prefetching workflow called Fluid, which drastically cut the load time for a 70B-class model from a multi-hour delay to just three minutes. The operational focus must therefore shift to optimizing the underlying model weight data access pipeline consistency.
Cut your AI search costs without sacrificing quality (The New Stack) — The New Stack
Vespa AI and Voyage AI introduce asymmetric retrieval to curb the operational expenditure of query-based AI search. This pattern bifurcates the embedding process: a high-fidelity model is reserved solely for creating the initial document index embeddings.
For every subsequent query, the system employs a much smaller, local model for the query embedding. This structural shift significantly minimizes query embedding costs, calculating a potential reduction from an estimated \$15,500/month to near zero at high query throughput (10k QPS).
CI wasn’t built for coding agents. Here’s what comes next. (The New Stack) — The New Stack
Traditional CI testing cycles are incompatible with the sub-second iteration loops inherent in coding agents. The feedback loop latency of minutes renders standard CI mechanisms unusable for immediate agent refinement.
The proposed solution is "plans"—mini, agent-authored end-to-end validation checks. These are designed to execute within the agent's live session against an integration environment, creating a validated artifact small enough to be instantly invoked by local tooling while still satisfying the fidelity requirements of a remote pipeline.
Building the agentic agreement enterprise: How developers are unlocking agentic experiences with Docusign’s MCP server and platform (Docusign MCP) — The New Stack
Docusign developed the Model Context Protocol (MCP) Server to bridge generalist LLMs with proprietary business logic in its Intelligent Agreement Management (IAM) platform. This directly addresses the gap between theoretical knowledge and technical capability.
The MCP Server functions by explicitly packaging and surfacing Docusign's domain-specific APIs and rules so that external LLM agents (like Claude or Gemini) can issue structured calls against the platform, enabling action rather than mere retrieval.
Datasette Agent (Simon Willison) — Simon Willison
The Datasette Agent provides natural language access to structured data within Datasette. A user query like "when was the last pelican sighting?" triggers the agent to generate and execute the necessary, complex SQL statement (e.g., SELECT title, commentary, created FROM blog_beat WHERE beat_type = 'sighting' AND (title LIKE '%pelican%' OR commentary LIKE '%pelican%') ORDER BY created DESC LIMIT 5).
This demonstrates a seamless capability for LLMs to generate verified, executable database queries that function as a true abstraction layer over complex SQL syntax.
datasette-agent-charts 0.1a2 (Simon Willison) — Simon Willison
The addition of "View SQL query" buttons below rendered datasette charts exposes the exact SQL query underpinning the visualization. This feature is vital for debugging and verifying data provenance.
It allows engineers to instantly inspect the precise transformation logic applied to derive any metric, providing a crucial, auditable step in any automated ML data pipeline.
datasette-agent-sprites 0.1a0 (Simon Willison) — Simon Willison
This plugin integrates the Datasette Agent with a Fly Sprites sandbox environment. Its function is to allow the agent to predictably execute external or specialized terminal commands from within its operational context.
This mechanism provides a necessary containerization and sandboxing layer, ensuring that while the agent can interface with external, specialized tools, its core environment remains isolated and manageable.
How to build your first end-to-end AI workflow in n8n (n8n) — The New Stack
The n8n article details constructing a complex, multi-stage AI content publishing pipeline. The workflow's sequence begins with triggering off content submitted into a Google Doc, simulating the first stage of a business process.
This implementation proves the viability of linking disparate services—Google Docs as a trigger, conditional routing logic, and AI nodes for content enrichment—into a single, reliable, and automated state machine flow.
Researcher: gemma4:e4b • Writer: gemma4:e4b • Editor: gemma4:e4b