AI Agent Ecosystem

Specialized intelligence, with capability and autonomy kept separate.

KAI includes real domain systems and a separate registry of 13 declared technical agent roles. The product systems are shown first. The registry below then exposes two statuses for each declared role, on purpose: Capability Status -- is the underlying work actually happening in production today -- and Autonomous Agent Status -- is it running as the declared standalone agent yet. Several capabilities run as pipeline stages or bounded orchestration adapters, while the 13 SubAgentSkeleton roles remain scaffolding.

13 agents cross-checked against the declared registry at the last data sync — 5 Operational, 7 Planned, and 1 Coming Soon. Generated from the KAI OS repository during the latest website data-sync run. Not live.

Operational System Map

The intelligence KAI can actually coordinate.

These are real product-level systems, distinct from the technical registry below. Their status reflects bounded capability today—not a claim of unrestricted autonomy.

Conversation Layer

Operational

Local conversation, multilingual routing, and spoken output with an offline-first fallback path.

See verified capabilities →

Content Production

Operational

A linear production system with real media stages and mandatory human review before publishing.

See the content flow →

Knowledge Brain

Operational

A callable, local keyword-search capability over the engineering report corpus; not a semantic universal memory.

Open Knowledge Brain →

Developer Memory

Operational

Append-only records for decisions, fixes, failures, and lessons, exposed through a bounded agent adapter.

See verified capabilities →

Trading Intelligence

In Development

Professional research, strategy, broker, and risk intelligence under controlled paper-only validation.

Review the paper-only boundary →

Desktop Operator

In Development

Permission and approval architecture for future controlled desktop actions; unrestricted control is not active.

Review the control architecture →

Intelligence

1 agent

CapabilityPlanned

Autonomous AgentPlanned

Research Agent

Would gather topic and trend research for future content ideas, with a stated evidence basis for every suggestion.

Inputs: A subject area, or a gap analysis against the existing topic queue.
Outputs: Candidate new topics, never presented with fabricated trend claims.
Dependencies: An unimplemented trend-intelligence subsystem (see Trend Agent).
Current Capability: Not built. Depends on trend-intelligence infrastructure that doesn't exist yet.
Automation Level: None.
Roadmap: Build after the Trend Agent's underlying data sources exist.

Last updated: 2026-07-31

Creation

4 agents

CapabilityOperational

Autonomous AgentIn Development

Visual Agent

Sources and manages scene visuals through the Visual Asset Ledger, with duplicate-checking against every prior video.

Inputs: Scene descriptions, a topic ID, a video ID.
Outputs: Sourced visual assets recorded in the ledger, with duplicate-check results per scene.
Dependencies: visual_intelligence/ (registry, ledger, duplicate_checker), Pexels video/photo API.
Current Capability: Runs automatically on every real video produced today, as inline pipeline logic.
Automation Level: Fully automated -- no manual step required.
Roadmap: Package this existing logic into a standalone, independently callable agent.

Last updated: 2026-07-31

See this in the pipeline flow →

CapabilityOperational

Autonomous AgentIn Development

Thumbnail Agent

Generates and ranks thumbnail candidates, avoiding backgrounds already used on prior videos.

Inputs: A video title/topic, target headline text options.
Outputs: Candidate thumbnails recorded in the ledger; final selection is human-reviewed.
Dependencies: visual_intelligence/thumbnail_memory.py, Pexels photo API.
Current Capability: Thumbnail generation and duplicate-avoidance already run automatically on every real video.
Automation Level: Fully automated for generation; final selection is human-reviewed.
Roadmap: Package into a standalone agent; add real CTR-prediction scoring (needs a vision model, not built yet).

Last updated: 2026-07-31

See this in the pipeline flow →

CapabilityOperational

Autonomous AgentIn Development

Music Agent

Selects background music by mood and checks reuse frequency before sourcing a track.

Inputs: A target mood tag, a video ID.
Outputs: A sourced track recorded in the ledger, or a reuse-frequency warning.
Dependencies: music_provider_router.py, Jamendo API / local cache.
Current Capability: Music sourcing and reuse-tracking already run automatically on every real video.
Automation Level: Fully automated.
Roadmap: Package into a standalone agent; add real mood-tag rotation (the same mood is currently reused often).

Last updated: 2026-07-31

See this in the pipeline flow →

CapabilityOperational

Autonomous AgentIn Development

Voice Agent

Synthesizes narration audio and records it in the asset ledger.

Inputs: Narration text, a topic ID.
Outputs: Synthesized narration recorded in the ledger.
Dependencies: narration_gateway.py, edge-tts.
Current Capability: Narration generation already runs automatically on every real video, currently limited to a single voice and provider.
Automation Level: Fully automated.
Roadmap: Package into a standalone agent; add a second voice/provider to choose between.

Last updated: 2026-07-31

See this in the pipeline flow →

Quality & Safety

1 agent

CapabilityOperational

Autonomous AgentIn Development

Compliance Agent

Reviews scripts and assets against compliance and licensing checks before anything is marked ready for review.

Inputs: A script/narration text, scene results, licensing metadata.
Outputs: A compliance report flagging any issues -- never an auto-publish signal.
Dependencies: compliance_gateway.py, fact_verification_gateway.py.
Current Capability: This exact check already runs automatically on every real video.
Automation Level: Fully automated for the check itself; publishing always requires human approval.
Roadmap: Package into a standalone callable agent rather than inline pipeline logic.

Last updated: 2026-07-31

See this in the pipeline flow →

Knowledge

2 agents

CapabilityPlanned

Autonomous AgentPlanned

Knowledge Agent

Searches production, CEO, engineering, and compliance reports to answer questions like "have we used this before?"

Inputs: A natural-language-ish query string.
Outputs: Matching report excerpts with file paths (keyword-matched, not semantic search).
Dependencies: knowledge_brain/.
Current Capability: The underlying search tool is real and works when run directly -- it isn't wired into the automatic production pipeline yet.
Automation Level: Not yet automated -- manual invocation only.
Roadmap: Wire into the production pipeline so it runs automatically; upgrade from keyword matching to semantic search.

Last updated: 2026-07-31

CapabilityPlanned

Autonomous AgentPlanned

Memory Agent

Records bugs, fixes, decisions, and lessons into a permanent, append-only ledger -- the "remember what happened" counterpart to every other agent.

Inputs: An event description (bug, fix, decision, asset generated).
Outputs: An appended, immutable record in the relevant ledger.
Dependencies: developer_memory/, visual_intelligence/ledger.py.
Current Capability: The ledger itself is real and actively used -- entries are written directly by the engineering process today, not by an autonomous agent.
Automation Level: Not yet automated -- manual writes.
Roadmap: Automate ledger writes as a direct side effect of other agents' actions, once those agents exist.

Last updated: 2026-07-31

Analytics

2 agents

CapabilityPlanned

Autonomous AgentPlanned

Analytics Agent

Would orchestrate real analytics collection into a summarized performance digest.

Inputs: A time window or list of topic IDs.
Outputs: A performance report, honestly scoped to whatever real data is actually available.
Dependencies: collect_analytics.py, YouTube Analytics API (OAuth scope not yet authorized).
Current Capability: Not running. Supporting code exists, but real analytics access isn't authorized or connected yet.
Automation Level: None.
Roadmap: Authorize the YouTube Analytics OAuth scope, then wire collect_analytics.py into a scheduled run.

Last updated: 2026-07-31

CapabilityComing Soon

Autonomous AgentComing Soon

Trend Agent

Would populate trend-monitoring data once real analytics and social monitoring exist.

Inputs: Not yet defined.
Outputs: Not yet defined.
Dependencies: Real analytics access and social-monitoring infrastructure -- neither exists yet.
Current Capability: Zero implementation. This is the one agent with no supporting code built at all.
Automation Level: None.
Roadmap: Build after the Analytics Agent's data access is authorized and connected.

Last updated: 2026-07-31

Operations

2 agents

CapabilityPlanned

Autonomous AgentPlanned

SEO Agent

Would review and optimize titles, descriptions, and tags against known-performing patterns.

Inputs: Draft metadata (titles, description, tags) for a video.
Outputs: Suggested metadata revisions with a stated rationale -- never auto-applied.
Dependencies: knowledge_brain/, content_intelligence.py -- both need real performance data neither has yet.
Current Capability: Not built. Needs enough real performance data to learn from, which doesn't exist yet.
Automation Level: None.
Roadmap: Build once enough real published-video performance data exists to learn from.

Last updated: 2026-07-31

CapabilityPlanned

Autonomous AgentPlanned

Website Agent

Would implement the defined website-integration interfaces once they're needed.

Inputs: Not yet defined beyond the interface layer.
Outputs: Not yet defined beyond the interface layer.
Dependencies: website_integration/ (interfaces only, no implementation).
Current Capability: This website is built today through direct human-AI collaboration, not by this agent. Only the interface layer exists in code.
Automation Level: Manual (human-AI collaboration).
Roadmap: Implement once there's a concrete, bounded website task worth automating end-to-end.

Last updated: 2026-07-31

Developer Infrastructure

1 agent

CapabilityPlanned

Autonomous AgentPlanned

Developer Agent

Analyzes code, finds bottlenecks, and proposes changes for human review -- sandbox, test, benchmark, approve, then production, never auto-deploy.

Inputs: A task description, relevant file paths, prior lessons learned.
Outputs: A proposed diff/plan and benchmark results, submitted for human approval.
Dependencies: developer_memory/ (for lesson lookup).
Current Capability: This work happens today through direct human-AI collaboration each engineering session, not an autonomous agent.
Automation Level: Manual (human-AI collaboration).
Roadmap: Define a bounded, sandboxed automation path once changes can be safely tested and benchmarked without a human in the loop for every step.

Last updated: 2026-07-31