Hermes puts multi-agent work on a board
Hermes Agent v0.12.0 added a Kanban-based multi-agent system. Agents claim tasks, work in parallel, and hand work off when they get blocked.
In the v0.12.0 announcement, the board keeps coordination state outside any one agent’s context. Each piece of work has an owner and status, while blocking relationships stay visible to agents and people. Later, Hermes also added orchestrator-driven auto-decomposition for triage tasks.
Flue makes the agent runtime programmable
Fred Schott introduced Flue, a TypeScript framework with a built-in agent runtime that works as a headless and programmable counterpart to an interactive coding agent.
The headless interface gives applications programmatic control over agent state, events, tools, and outputs. Flue leaves choices such as tool permissions, event handling, and state ownership visible to the application rather than hiding them behind an interactive session.
Delegation as something models can learn
Research on training models to use sub-agents teaches a model when to hand work to sub-agents, how to divide the problem, and how to combine their results.
A lot of multi-agent systems quietly assume a model already knows when to delegate and how to integrate the result. This work treats those decisions as things to train and evaluate. Models can split a task badly, duplicate effort, or accept a sub-agent’s answer without checking it.
Training may improve those choices, but the test also has to count the extra model calls and coordination time. Otherwise a system can look better by spending much more compute on every problem.
Runnable workshop examples
An AI Engineer workshop was open-sourced with runnable implementations. The examples include a research agent using grounded search and YouTube analysis, plus a generate-review-edit workflow.
Runnable examples let readers inspect the retrieval, review, and handoff steps instead of inferring them from slides. They show where sources enter the system and how a draft changes.
Printing Press reconsiders the CLI
Printing Press is a CLI factory and library. Its authors argue that conventional APIs, MCP servers, and CLIs impose unnecessary token and interaction overhead on agents.
An agent may need a different interface from a person. Decorative output, interactive prompts, unstable formatting, and large schemas can consume context and make automation brittle. A compact, deterministic CLI offers a smaller action surface and gives the agent runtime cleaner evidence to record. The cost is another adapter layer that has to follow upstream API changes.
One repeated loop in HeavySkill
A post about HeavySkill argues that one reusable skill can improve agent results by running several attempts in parallel and then comparing them.
The proposed loop does this instead of arranging a larger graph of specialized roles. Whether that holds up depends on the task and cost budget. Independent attempts help some work; sequential work may gain very little from running more attempts.
Using existing agent runtimes for training
Polar runs training experiments using existing Codex, Claude Code, OpenClaw, Hermes, or custom agent runtimes.
Those systems already contain tools, execution environments, and task structure. Polar uses them directly to produce training runs instead of rebuilding a simplified version of each agent stack.