May 2026 Summaries

Conductor-skills is an AI knowledge package designed to enable AI coding agents to efficiently create, run, monitor, and manage Conductor workflows without requiring explicit user instructions. It addresses the challenge of agents providing inaccurate or incomplete guidance by offering a structured set of files that the AI reads to gain a comprehensive understanding of Conductor. This includes commands, workflow JSON schemas, and error handling for various task types. The core of the skill is a structured Markdown document that provides rules, setup procedures, and command references, ensuring that agents can autonomously handle tasks like server connections, workflow execution, and monitoring. Additionally, the skill includes reference files and examples to guide agents through complex operations, and it provides a fallback Python script for environments where the Conductor CLI is unavailable. By packaging this knowledge in a language-agnostic format, Conductor-skills can be utilized across various AI platforms, enhancing workflow management while minimizing errors and inefficiencies.

AI agents lack persistent recall unless memory is explicitly designed, making semantic memory essential for retaining durable facts, preferences, decisions, and domain knowledge across separate runs. The text distinguishes semantic memory from conversation memory, which preserves recent messages, and workflow state, which records execution details such as tool results, approvals, and failures. It presents Agentspan as a framework that separates these forms of state through components for chat history, long-term knowledge retrieval, and server-side execution tracking. A recommended implementation exposes semantic memory as a tool, allowing agents to decide when relevant context is needed while enabling applications to govern retrieval and access. Developers can use the same memory interface for temporary local testing or durable production backends such as Pinecone, Weaviate, ChromaDB, Qdrant, or Mem0, with backend systems managing persistence, relevance, tenant isolation, expiration, and deletion.

Agentic AI is often promoted as a broad marker of modernization, but production-ready agents should be evaluated by their ability to reliably coordinate models, tools, state, policies, human approvals, and failure recovery rather than simply complete isolated tasks. These systems are best understood as distributed execution workflows in which language models guide decisions while runtime infrastructure manages persistence, retries, idempotency, observability, auditing, and safe resumption after process or API failures. Integrations create much of the operational risk because external systems differ in reliability and consequences, requiring clear sources of truth, bounded API access, credential management, retry rules, and approval controls for high-impact actions such as billing or deletion. Tool design should distinguish between low-risk reads, retry-safe writes using idempotency keys, and sensitive operations that require human authorization, while multi-agent workflows use sequencing, parallel checks, routing, and manual intervention to constrain how work proceeds. A durable deployment architecture includes tracked entry points, a coordinating runtime, persistent execution state, replaceable workers, approval layers, and logs, metrics, traces, and execution histories. Agentspan is presented as an open-source distributed runtime intended to provide these capabilities, including crash recovery, distributed tools, durable approval pauses, monitoring, and support for agents built with several existing frameworks.

Prompt Studio, part of Orkes Conductor, is a dedicated environment designed for engineering and testing AI prompts in isolation, allowing for rapid iteration without the need to execute the full workflow. It enables users to parameterize prompts with variable substitution and compare outputs across different large language models (LLMs), facilitating informed decisions on model selection without being confined to a single provider. The platform supports testing prompts on various scenarios, such as release note generation from PR descriptions, invoice field extraction, SQL generation, and meeting transcript analysis, ensuring prompts are robust and effective before deployment. By allowing users to bring their preferred LLMs and easily integrate them, Prompt Studio provides flexibility and cost efficiency, making it a valuable tool for refining AI prompts before integrating them into workflows.

Agentspan is an orchestration platform for building, importing, and observing durable AI agent workflows, supporting native SDK agents as well as integrations with LangGraph, the OpenAI Agents SDK, and Google ADK. Its `run()` method executes an agent synchronously and returns only a final result, while `stream()` yields real-time events for reasoning, tool calls, tool results, and completion, making it useful for debugging, progress interfaces, monitoring, and human-in-the-loop workflows. An operations-agent example shows how streaming exposes an investigation from service health checks through database-log analysis to a targeted service restart, whereas the final outcome from `run()` omits those intermediate actions. Agentspan also offers `start()` for asynchronous execution with later event consumption through a handle, which suits decoupled HTTP and WebSocket architectures. Streaming does not alter workflow durability because execution continues server-side if a client disconnects, although reconnecting resumes only current and future events rather than replaying prior ones; agents using decorated tools must also be registered with `runtime.serve()` so tool calls can be dispatched.

The article provides a comprehensive guide to building AI agents using Orkes Conductor, a workflow orchestrator, and Agentspan, an open-source agent builder and runtime. It explains the concept of AI agents, which are software programs that use large or small language models to execute tasks based on given goals, utilizing tools, loops, and memory for enhanced functionality. The guide walks through the development of a subscription analyzer agent that can search email inboxes for recurring charges, flag unused subscriptions, and suggest potential savings. It emphasizes the ease of using Agentspan for defining and running AI agents by abstracting the orchestration complexities handled by Conductor, allowing users to integrate agents into existing workflows. The article also highlights the flexibility of agent definitions, which can be transformed into different applications by simply swapping tools and instructions, and offers detailed steps for connecting the agent to a real Gmail inbox for practical use.

Agentspan’s router strategy separates task classification from specialist execution by using a dedicated router, such as a lower-cost LLM or deterministic Python function, to select the appropriate agent before work begins. The example issue-triage workflow routes bug reports, feature requests, and documentation questions to specialized GPT-4o handlers while a GPT-4o-mini classifier returns only the relevant agent name, reducing costs and making routing explicit compared with the handoff strategy, where a parent model both interprets and delegates tasks. Agentspan supports durable server-side workflows that preserve routing decisions across crashes, provides execution observability through a dashboard, and allows router-based workflows to be combined with sequential pipelines or nested research processes. The strategy concludes an eight-part overview of Agentspan coordination patterns, presenting routing as a configurable option within the same Agent class rather than a separate framework.

Agentspan’s manual multi-agent strategy places humans directly in control of workflow routing by pausing after each turn and requiring a person to choose which agent acts next. The post demonstrates this approach through an editorial team of a writer, fact checker, and copy editor, where a human can dynamically direct drafting, verification, revision, and polishing according to the content’s needs rather than a fixed sequence or automated decision. Using the same Agent class as Agentspan’s other orchestration strategies, a team configured with `Strategy.MANUAL` emits a waiting event, accepts a selected agent through a response call, and repeats until reaching its turn limit. The strategy is positioned for editorial review, education, debugging, and quality-control workflows where human judgment is the routing logic, while automated strategies remain preferable when rules, model decisions, or fixed ordering are sufficient. Agentspan preserves paused workflows on its server indefinitely, allowing users to reconnect or respond later without blocking a process.

Agentspan’s random multi-agent strategy selects one agent at random on each turn, allowing the same participant to speak repeatedly or be skipped, unlike round-robin coordination that guarantees a fixed, equal rotation. The approach is presented as particularly suitable for brainstorming and creative exploration, where unpredictable contributions from creative, practical, and critical thinkers can produce varied perspectives before a summarizer identifies leading ideas, risks, and next steps. Agentspan serves as a durable orchestration layer that can build or integrate AI agents, persist workflows through failures, and expose execution details in a dashboard. A random strategy can run for a defined number of turns, retain completed selections after a crash, and be combined with other strategies, such as using random ideation followed by structured round-robin review. It is recommended when variety, load balancing, stress testing, or collaborative creative output is more important than ensuring every agent participates equally.