AI Engineer · Published 2026-09-27

We're the bottleneck, but we don't have to be — Ido Salomon, AgentCraft

privateReview: unreviewedConfidence: high
agentcraftorchestrationmulti-agentAutonomyvisibilityrtsagent-orchestrationHuman-in-the-loopattention-managementsituational-awarenessreview-artifactsagent-federationgame-inspired-ui
Open on YouTube ↗

Summary

Overview

  • Speaker: Ido Salomon
  • Channel: AI Engineer
  • Main topic: Agent orchestration, visibility, autonomy, and collaboration in AI software engineering
  • Purpose: Explain how agent orchestration can shift humans from being the execution bottleneck to collaborative directors of multi-agent AI systems. Ido Salomon, creator of AgentCraft and MCP-UI, discusses the bottlenecks of working with AI agents at scale. While agents are powerful, developers face cognitive overload from managing multiple agents, frequent baby-sitting, and context switching. To solve this, AgentCraft introduces an RTS-game-inspired orchestrator that brings visibility, autonomy, and collaboration into agent management, enabling multi-agent workflows and lowering the barrier for non-technical users.

Topic Map

The Agent Scaling Bottleneck

  • Explanation: Discusses why developers aren't all unstoppable even though agents are powerful: managing multiple cloud agents creates massive cognitive overhead, fatigue, and burnout.
  • Key claims:
    • Agents are amazing for tasks from cat memes to SaaS apps
    • Spawning 25 agents is easy, but steering, directing, and reviewing them at scale is exhausting
    • Humans are currently the bottleneck in multi-agent workflows
  • Examples:
    • Spinning up 25 cloud code agents on multiple monitors
  • Terminology:
    • AgentCraft
    • cloud code agents
    • orchestrator
  • Why it matters: Demonstrates that managing AI agent scale requires more than just compute and model access—it requires better UX and orchestration.

AgentCraft: Game-Inspired Orchestration

  • Explanation: Introduces AgentCraft, an orchestrator inspired by real-time strategy (RTS) games like Warcraft and Sims to manage AI agents visually.
  • Key claims:
    • Managing AI agents resembles managing units in a real-time strategy game
    • AgentCraft provides visual maps, buildings representing functionalities, and side panels for prompts and voice
    • Brings video game interaction models into developer tooling
  • Examples:
    • Warcraft and Sims unit management analogies
    • AgentCraft side panel with voice and text prompting
  • Terminology:
    • AgentCraft
    • RTS orchestrator
    • multimodality
  • Why it matters: Provides a familiar paradigm for supervising complex distributed systems of AI agents.

Raising the Ceiling: Visibility and Autonomy

  • Explanation: Explores how providing visibility into agent file systems and tasks, combined with autonomous loops, raises the ceiling for power users.
  • Key claims:
    • Visibility lets developers see what agents are doing on a map representing file systems
    • Autonomy allows agents to accept quests and execute tasks without constant babysitting
    • Review kits provide diffs, file changes, and visual evidence
  • Examples:
    • Mapping file system directories to map ruins
    • Spacebar jump to nearest agent needing attention
    • Software Factory autonomous execution
  • Terminology:
    • visibility
    • autonomy
    • review kit
    • software factory
  • Why it matters: Enables power users to manage dozens of concurrent tasks efficiently without getting overwhelmed.

Lowering the Floor: Collaboration and Community

  • Explanation: Addresses how team collaboration features ('War Halls') and simplified mobile interfaces lower the floor, allowing non-technical users and broader communities to leverage AI agents.
  • Key claims:
    • Collaboration allows multiple team members and agents to share workspaces and communicate via notice boards
    • War halls enable multi-machine agent federation
    • Lowering the floor brings non-technical users into agent orchestration
  • Examples:
    • War Halls for team product designers and engineers
    • Mobile device interaction with AgentCraft
    • Feedback from kids and gamers using agent orchestration
  • Terminology:
    • War Halls
    • collaboration
    • notice board
    • mobile level
  • Why it matters: Expands the utility of AI agents beyond hardcore developers to non-technical users and teams.

Key Points

Humans are the bottleneck in multi-agent systems

  • Explanation: While spawning multiple AI agents is trivial, steering and reviewing them at scale causes severe cognitive fatigue.
  • Evidence: Screenshots of developers monitoring dozens of active agent chat windows.
  • Practical implication: Tooling must evolve from simple chat interfaces to sophisticated orchestrators with clear visibility and management UX.

RTS game mechanics translate naturally to agent orchestration

  • Explanation: Managing units in strategy games like Warcraft involves spatial awareness, task assignment, and supervision, mirroring multi-agent development.
  • Evidence: AgentCraft interface displaying agents as units, buildings as plugin/skill managers, and mini-maps for file tracking.
  • Practical implication: Adopting game design patterns can dramatically improve human-AI interaction and task management efficiency.

Balancing autonomy with visibility and collaboration

  • Explanation: True productivity gains require both high ceilings for power users through autonomy and low floors for non-technical users through simplified collaboration tools.
  • Evidence: User feedback from gamers and non-developers successfully orchestrating agents using game-like interfaces.
  • Practical implication: AI engineering tools should prioritize collaboration spaces and intuitive interfaces alongside raw model capability.

Frameworks, Models & Processes

AgentCraft Orchestration Framework

  • How it works: Combines game-inspired spatial maps with agent monitoring, autonomous task execution loops, and team collaboration rooms.
  • Components:
    • Visibility map and file system projection
    • Autonomous task execution and review kits
    • War Halls for team and multi-agent collaboration
  • When to use: When managing multiple concurrent AI coding and task execution agents.

Examples & Case Studies

Ido Salomon built AgentCraft to mirror real-time strategy games for agent management.

  • Illustrates: How game design patterns can solve developer cognitive overload in AI agent workflows.
  • Lesson: Familiar UI paradigms from gaming can make complex multi-agent orchestration intuitive.

Actionable Takeaways

  • Immediate:
    • Check out AgentCraft via npm install
    • Explore visual orchestration for multi-agent workflows
  • Strategic:
    • Shift human roles from execution bottleneck to collaborative directors
    • Design AI engineering tools with both high ceilings for power users and low floors for non-technical participants
  • Questions to investigate:
    • How can multi-agent collaboration scale across distributed teams?
    • What other gaming metaphors can improve AI agent supervision?

Claims Worth Verifying

  • AgentCraft is available via npm install (technical)

Notable Quotes

"We're the bottleneck, but we don't have to be" "Agents are magic" "We are the bottleneck"

Compressed Summary

  • Managing multiple AI agents causes developer burnout and cognitive overload.
  • AgentCraft applies RTS video game mechanics to multi-agent orchestration.
  • Visibility, autonomy, and collaboration successfully raise the engineering ceiling.
  • Game-inspired interfaces also lower the floor for non-technical users.
  • Keywords: agentcraft, orchestration, multi-agent, autonomy, visibility, rts
  • Core insight: Applying real-time strategy game mechanics to AI agent orchestration solves human cognitive bottlenecks and enables seamless multi-agent collaboration.

Core insights

6
Mental Modelmedium noveltymoderate evidence

The binding constraint in multi-agent systems is human attention bandwidth, not agent supply: spawning many cloud agents is trivial, but steering, directing, and reviewing them is what causes fatigue.

Why it matters

It relocates the optimization target of multi-agent architecture from throughput (more agents, more parallelism) to supervisory load per unit of agent output. Any design that increases concurrent agents without reducing per-agent review cost is net-negative.

Generalization

Scaling any autonomous-worker system is bounded by the supervisor's ability to triage, direct, and verify — so engineering effort should shift to attention routing, verification artifacts, and exception surfacing rather than to raw concurrency.

Spawning 25 agents is easy, but steering, directing, and reviewing them at scale is exhausting
Open source video
Humans are currently the bottleneck in multi-agent workflows
Open source video
Architecturehigh noveltymoderate evidence

Spatial RTS-style UI is proposed as a concrete orchestration abstraction: agents become units on a map, the file system becomes terrain ('map ruins'), and plugins/skills become buildings the user places.

Why it matters

This is a claim that agent state should be projected into a persistent spatial workspace rather than into N independent chat transcripts. It changes the primary UI primitive from conversational stream to navigable world-state, which is what makes supervising dozens of concurrent entities tractable.

Generalization

For any system with many concurrent, long-running autonomous entities, a shared spatial/state projection can replace per-entity transcript monitoring as the human interface layer.

Managing AI agents resembles managing units in a real-time strategy game
Open source video
Mapping file system directories to map ruins
Open source video
buildings as plugin/skill managers, and mini-maps for file tracking
Open source video
Mechanismhigh noveltyweak evidence

Attention routing is treated as a first-class orchestration primitive: a hotkey jumps the operator to the nearest agent needing attention, instead of the operator polling many windows.

Why it matters

It implies the orchestrator maintains a notion of 'needs human attention' as system state and pushes the human toward it — an interrupt/triage model rather than a monitoring dashboard. This is the difference between O(N) scanning and O(k) focused review where k is the number of blocked agents.

Generalization

Agent harnesses should expose an explicit 'blocked / needs-decision' state per agent and an O(1)-ish path to it, decoupling human cost from total agent count.

Spacebar jump to nearest agent needing attention
Open source video
Visibility lets developers see what agents are doing on a map representing file systems
Open source video
Practicemedium noveltymoderate evidence

Review artifacts (diffs, file changes, visual evidence) are framed as the productized output of an autonomous loop — the human consumes a review kit, not the agent's chat log.

Why it matters

It decouples agent execution length from human attention cost: the agent can run an autonomous loop as long as its termination state is a compact, verifiable artifact. The review interface becomes the real contract between agent runtime and operator.

Generalization

Autonomy is only safe to grant in proportion to the quality of the synthesized evidence bundle produced at handoff; the harness should emit evidence as a first-class artifact, not as a byproduct of conversation.

Autonomy allows agents to accept quests and execute tasks without constant babysitting
Open source video
Review kits provide diffs, file changes, and visual evidence
Open source video
Tradeoffmedium noveltyweak evidence

Orchestration tooling is designed along two independent axes — 'raise the ceiling' (visibility + autonomy for power users) and 'lower the floor' (simplified/mobile/collaborative entry points) — rather than as one capability ladder.

Why it matters

It asserts a single interface cannot serve both expert supervisors of dozens of agents and non-technical participants, so capabilities must be layered rather than averaged. This constrains product and API surface decisions in agent platforms.

Generalization

Agent platforms should expose an expert control plane and a separate low-friction participant plane sharing the same underlying state, instead of one generalized UI.

Lowering the floor brings non-technical users into agent orchestration
Open source video
True productivity gains require both high ceilings for power users through autonomy and low floors for non-technical users through simplified collaboration tools
Open source video
Architecturemedium noveltyweak evidence

Collaboration is modeled as shared-space federation: humans and agents share workspaces and a notice board, and 'War Halls' federate agents across multiple machines.

Why it matters

It treats agents as peer participants in a shared communication substrate that spans hosts, rather than as processes owned by one developer's client. That is an architectural commitment to cross-machine agent addressing and durable shared context.

Generalization

Multi-agent systems that span people and machines need a shared, addressable message/workspace layer (board + room abstraction), not point-to-point human-to-agent chat.

Collaboration allows multiple team members and agents to share workspaces and communicate via notice boards
Open source video
War halls enable multi-machine agent federation
Open source video

Deep dives

5

Attention cost per completed agent task as the real scaling metric

Research question

Does autonomous execution plus a review-artifact handoff actually reduce total human attention minutes and interventions per completed task compared to turn-by-turn babysitting, or does it merely relocate the cost from mid-run steering to end-of-run review?

Why

If the binding constraint is supervisory attention, then every orchestration feature should be evaluated by human-minutes-per-task and defect-escape-rate, not by agent count or parallelism. Without that measurement, 'autonomy' features are unfalsifiable product claims.

Spawning 25 agents is easy, but steering, directing, and reviewing them at scale is exhausting
Open source video
Humans are currently the bottleneck in multi-agent workflows
Open source video
Autonomy allows agents to accept quests and execute tasks without constant babysitting
Open source video
Source video

Detecting and exposing 'needs human decision' as a first-class agent state

Research question

Can an agent runtime classify its own blocked/uncertain states accurately enough that a jump-to-nearest-agent-needing-attention affordance yields O(k) human review (k = blocked agents) rather than O(N) polling, and what are the false-positive/false-negative rates versus human judgment?

Why

Attention routing only pays off if the signal that drives it is trustworthy. If the orchestrator over-flags, the human scans anyway; if it under-flags, blocked agents stall silently. This makes blocked-state classification the load-bearing component of the whole triage model.

Spacebar jump to nearest agent needing attention
Open source video
Visibility lets developers see what agents are doing on a map representing file systems
Open source video
Source video

Evidence-bundle sufficiency: what a review kit must contain to license autonomy

Research question

Which classes of agent error are surfaced by a diff/file-change/visual-evidence review kit that a conversational transcript would miss, and which errors escape both — i.e. what is the minimum sufficient evidence bundle for a given autonomy horizon?

Why

Granting longer unsupervised runs converts continuous supervision into batched, deferred review. The failure mode shifts from mid-run drift to invisible post-hoc approval, so the composition of the handoff artifact becomes the actual contract between agent runtime and operator.

Review kits provide diffs, file changes, and visual evidence
Open source video
Autonomy allows agents to accept quests and execute tasks without constant babysitting
Open source video
Source video

Legibility limits of spatial state projection under rising concurrency

Research question

At what agent count does a shared spatial/world-state projection stop improving situational awareness over a multi-window transcript baseline, measured by time-to-detect-anomaly and time-to-assign-task across agent counts (e.g. 5/15/40)?

Why

The RTS map is claimed to make global state cheap to acquire, but no scaling curve is reported. If there is a ceiling, orchestration architecture needs hierarchical aggregation (regions, mini-maps) as a designed requirement rather than an aesthetic choice.

Managing AI agents resembles managing units in a real-time strategy game
Open source video
Mapping file system directories to map ruins
Open source video
Brings video game interaction models into developer tooling
Open source video
Source video

Two-plane orchestration architecture: expert control plane vs participant plane over shared state

Research question

Can an expert control plane (visibility + autonomy for supervisors of dozens of agents) and a low-friction participant plane (simplified, mobile, collaborative entry points) share one underlying state model without the simplification of the participant plane degrading the verification guarantees of the expert plane?

Why

If the two audiences genuinely require divergent interfaces, capability layering constrains API and product surface decisions in agent platforms — and it predicts a specific failure: participants approving outputs they cannot audit.

Lowering the floor brings non-technical users into agent orchestration
Open source video
True productivity gains require both high ceilings for power users through autonomy and low floors for non-technical users through simplified collaboration tools
Open source video
Source video

Article ideas

4

Your multi-agent system's real bottleneck is a queue of one

Adding concurrent agents without reducing per-agent review cost is net-negative: the supervisor is a single-server queue, so throughput gains only materialize when the design reduces interventions per completed task — which means orchestrator engineering should be spent on triage, verification artifacts, and exception surfacing rather than on raw parallelism.

Angle

Reframe multi-agent scaling as a human-factors and queueing problem, then derive concrete design obligations (interrupt-driven oversight, first-class blocked state, evidence bundles) from that reframing.

Source video

Stop reading transcripts: the handoff artifact is the real interface

Autonomy is only safe to grant in proportion to the quality of the synthesized evidence bundle produced at handoff, so the review kit — diffs, file changes, visual evidence — should be engineered as a first-class product artifact rather than emitted as a byproduct of conversation.

Angle

Argue the contract between agent runtime and operator is the termination artifact, and show why bad evidence bundles become the dominant failure mode as run length grows.

Source video

The map is the interface: spatial world-state beats N chat windows

When many long-running autonomous entities run concurrently, projecting their state into a persistent, navigable spatial workspace makes situational awareness O(1) to acquire instead of O(N) transcript reads — but the approach only holds if the spatial model stays faithfully synchronized with real system state.

Angle

Treat the map as a state-visualization layer analogous to dashboards over raw logs, and confront the distortion risk of the game metaphor directly.

Source video

Raise the ceiling, lower the floor — and stop pretending one UI does both

A single generalized agent interface cannot serve both expert supervisors of dozens of agents and non-technical participants; platforms should expose an expert control plane and a separate low-friction participant plane over shared state, and must confront the verification gap that the low plane creates.

Angle

Argue against the one-capability-ladder assumption in agent product design, and name the specific risk: users approving diff-level output they cannot evaluate.

Source video

Project ideas

4

Attention Ledger

movement-lab

For the same task set, autonomous-run-plus-review-kit mode reduces human minutes per completed task and interventions per task versus turn-by-turn babysitting, without increasing defect-escape rate beyond a preset tolerance.

Proof of concept

Instrument a small harness that runs a fixed set of repo tasks in two modes (babysat vs. autonomous with an auto-generated diff/file/visual evidence bundle), logging every human action with timestamps and classifying each intervention.

Measurement

Human minutes per completed task, interventions per task, and post-hoc defect detection rate across both modes on the same task corpus.

Source video

Needs-Attention Signal Bench

beyond-evals

Per-agent blocked/uncertain state signals can be emitted with high enough precision that hotkey triage visits fewer agents than a polling baseline while surfacing a comparable share of genuinely blocked agents.

Proof of concept

Build per-agent state machines that emit an explicit 'needs human decision' status (with reason codes), then have operators triage a scripted multi-agent run via a jump-to-next-flagged affordance and label each flagged/ignored agent as truly blocked or not.

Measurement

Precision and recall of the needs-attention signal against human labels, plus flags-visited per genuinely blocked agent versus a poll-every-agent baseline.

Source video

Review Kit Ablation Study

beyond-evals

Evidence modality matters: a diff-only bundle misses a measurable class of errors caught by visual evidence, and a transcript-first review misses errors the diff surfaces — so no single modality is a sufficient handoff artifact.

Proof of concept

Inject a controlled set of seeded defects into agent-completed tasks, then present reviewers with three handoff conditions (transcript only, diff/file changes only, diff plus visual evidence) and record which defects they flag, with time on task.

Measurement

Per-modality defect detection rate by defect class, false-flag rate, and review time per task across the three conditions.

Source video

Map Legibility Probe

new

A spatial world-state projection yields faster time-to-detect-anomaly and time-to-assign-task than a multi-window transcript baseline up to some agent count, beyond which the advantage inverts without hierarchical aggregation.

Proof of concept

Build two supervisory front-ends over the same simulated fleet — a spatial map with regions and a mini-map, and a multi-window transcript view — and run operators through anomaly-detection and task-assignment scenarios at swept agent counts.

Measurement

Time-to-detect-anomaly and time-to-assign-task at each agent count, plus missed-anomaly rate, comparing map versus windowed baseline.

Source video

Architectural implications

4

The human interface to agents is reframed from a per-agent chat window to a map of world state.

Before

One tab or window per agent; the operator reconstructs global state by reading N transcripts.

After

A persistent spatial projection where each agent's location encodes what it is working on (e.g., directories as map regions) and a mini-map summarizes file/state activity.

Consequence

Global situational awareness becomes O(1) to acquire instead of O(N) reads, but the orchestrator must maintain a live bidirectional mapping between real system state and the spatial model — a nontrivial synchronization and consistency burden.

Source video

Autonomous execution loops are paired with mandatory review-artifact generation.

Before

Human babysits the agent turn-by-turn; review happens inside the conversation.

After

Agent accepts a task ('quest') and runs to completion, returning a diff/file-change/visual-evidence bundle for review.

Consequence

Shifts the harness requirement from 'streaming output' to 'synthesizing verifiable evidence', and shifts risk from mid-run drift to post-hoc review quality — bad evidence bundles become the dominant failure mode.

Source video

Orchestrator maintains an explicit attention/priority state per agent that the human navigates rather than polls.

Before

Operator scans all active agents to discover which is blocked or off-track.

After

System classifies agents as needing attention and provides a direct jump to the nearest such agent.

Consequence

Requires the runtime to detect and expose 'needs human decision' as a first-class signal — meaning the agent loop must recognize its own blocked states, which is a hard and underspecified classification problem.

Source video

Collaboration layer is separated from execution layer via shared rooms and notice boards spanning machines.

Before

Each developer's agents live in their own local environment; coordination is out-of-band.

After

Team members and agents share a workspace and a durable message board; agent federation spans multiple machines.

Consequence

Introduces distributed-systems concerns (identity, message ordering, conflicting concurrent work on shared state) into what was previously a single-user tool.

Source video

Tradeoffs and failure modes

4

Autonomy vs. review debt

Benefit

Agents accept tasks and execute without constant babysitting, raising the ceiling for power users.

Cost or risk

Autonomy converts continuous supervision into deferred, batched review; if the review kit is incomplete or unconvincing, errors propagate through longer chains before detection.

Autonomy allows agents to accept quests and execute tasks without constant babysitting
Open source video
Source video

Lowering the floor vs. verification rigor

Benefit

Non-technical users, designers, and even children can orchestrate agents through game-like, mobile-accessible interfaces.

Cost or risk

Users with less ability to audit diffs and file changes may approve agent output they cannot evaluate, weakening the human check that the review kit is meant to provide.

Lowering the floor brings non-technical users into agent orchestration
Open source video
Source video

Game metaphor as orchestration UX

Benefit

Familiar RTS conventions (units, buildings, mini-map) give users a ready-made mental model for supervising distributed agents.

Cost or risk

The metaphor may distort or hide the underlying model — users reason about 'units' and 'quests' rather than actual process, tool, and permission boundaries, and the mapping can drift from reality.

Brings video game interaction models into developer tooling
Open source video
Source video

Multi-machine agent federation

Benefit

War halls let multiple people and agents across machines share one workspace.

Cost or risk

Federated shared state introduces coordination, ordering, and conflict-resolution problems absent in single-machine orchestration.

War halls enable multi-machine agent federation
Open source video
Source video

Open questions

4

Does autonomous execution with review kits actually reduce total human attention per task, or does it merely shift cost from mid-run steering to end-of-run review?

Why unresolved

The summary asserts autonomy 'without constant babysitting' and lists review-kit contents but gives no measurement of review time, defect escape rate, or attention cost per completed task.

Research direction

Instrument a harness to log human minutes and number of interventions per completed task under babysitting vs. autonomous-plus-review-kit modes, and compare defect detection rates.

Source video

What signal does an orchestrator use to decide an agent 'needs attention' well enough to support hotkey triage across dozens of agents?

Why unresolved

The summary describes the UI affordance (spacebar to the nearest agent needing attention) but not how blocked/unproductive states are detected.

Research direction

Prototype per-agent state machines that emit explicit blocked/uncertain status, then measure false-positive and false-negative rates against human judgment.

Source video

Does spatial/RTS orchestration remain legible as concurrency grows, or does the map itself become an overload surface beyond some agent count?

Why unresolved

Claims rest on demonstrations with roughly 25 agents and anecdotal user feedback; no scaling curve or ceiling is reported.

Research direction

Run a supervised study sweeping agent count (e.g., 5/15/40) measuring time-to-detect-anomaly and time-to-assign-task across map UI vs. multi-window baseline.

Source video

How should shared workspaces with humans, agents, and notice boards resolve conflicting concurrent edits to the same files or artifacts?

Why unresolved

Collaboration and multi-machine federation are described as features, but no concurrency-control or conflict-resolution mechanism is given.

Research direction

Apply existing distributed-collaboration models (locking, CRDTs, branch-and-merge) to agent-authored changes and evaluate conflict frequency under realistic team workloads.

Source video

Key claims

7
causalVerification needed

Managing many agents at scale produces developer cognitive overload, fatigue, and burnout.

Evidence

Managing multiple AI agents causes developer burnout and cognitive overload.

Question

What measurable attention or error-rate degradation appears as concurrent agent count increases, and at what threshold?

Source video
opinionVerification needed

The human, not compute or model capability, is the current limiting factor in multi-agent workflows.

Evidence

We are the bottleneck

Question

Where do real multi-agent pipelines actually stall — on human review latency, on model quality, or on infrastructure limits?

Source video
comparativeVerification needed

RTS game mechanics map naturally onto multi-agent orchestration tasks (spatial awareness, unit tasking, supervision).

Evidence

Managing AI agents resembles managing units in a real-time strategy game

Question

Do users of game-inspired orchestration UIs outperform users of transcript-based UIs on supervision tasks?

Source video
factualVerification needed

AgentCraft projects the file system into the orchestration map.

Evidence

Mapping file system directories to map ruins

Question

How is the file-system-to-map projection maintained as directories change, and does it stay faithful under concurrent agent writes?

Source video
causalVerification needed

Autonomous loops can replace constant human babysitting when paired with review artifacts.

Evidence

Review kits provide diffs, file changes, and visual evidence

Question

What fraction of agent errors does a diff/file/visual review kit surface that a chat transcript would miss, and vice versa?

Source video
factualVerification needed

AgentCraft supports multi-machine agent federation through 'War Halls'.

Evidence

War halls enable multi-machine agent federation

Question

What is the coordination model (shared state, message bus, addressing) behind cross-machine federation, and how are conflicts handled?

Source video
factualVerification needed

AgentCraft is distributed as an npm package.

Evidence

AgentCraft is available via npm install (technical)

Question

Is the package publicly installable, maintained, and what is its license and runtime footprint?

Source video

Connections

4