Skip to content
The lab

An organization of
AI agents

Most talk about AI. We run a 51-agent organization in production — an executive board, teams and sub-agents — and expose both the architecture and the live gauges behind it.

The fleet, running

The same architecture, live in production.

Agents serving this site and the platform behind it — runs per minute, models in use, success rate and latency, with the infrastructure they run on.

Live · 9 agents active · synced 2 d ago
/lab · sden.ai
Agents in production
0

9 running right now

Runs · last 24h
0

~3.3/min sustained

Success rate
0.0%

Across all production agents

Latency p50
0 ms

p95 4200 ms

Orchestration

Agent fleet

Intake triage1,840 runsrunning
Site audit streamer1,120 runsrunning
Lead router760 runsrunning
Doc retriever (RAG)540 runsrunning
Prompt optimizer310 runsidle
Readiness scorer250 runsidle

Inference

Models in use

claude-haiku-4-571%
claude-sonnet-4-629%
Tokens · 24h18.4M in · 2.1M out

Capabilities

Tool calls · 24h

web_search9,120
retrieval6,340
code_exec2,110
send_email480
Est. compute cost$41.7
Platform health

The infrastructure the fleet runs on

Uptime
0.00%

Over the last 30 days · 0 incidents

Deploys this week
0

Last deploy 2 d ago · avg 17.3/week

Build time
0s

Avg 47s · p95 71s

Lighthouse perf
0

A11y 100 · SEO 100

Performance budget

Core Web Vitals

LCP1180 ms≤ 2500
INP96 ms≤ 200
CLS0.01≤ 0.1
JS bundle142 KBbudget 180 KB

Source control

Repositories

Monitored repos14
Open issues7
Avg PR review4.1 h
Top languagesTypeScript · Python · Rust

Reliability

Incidents

Open P10
Open P20
Last P1never
Last build failure1 mo ago

What's published

Catalog

Blog articles14
Expertise domains11
Products in production3
Case studies5

How this page works

A live window into the agents we run.

Fleet and platform figures are drawn from our own orchestration layer, git history, Lighthouse CI, and production monitoring. The page is statically exported and animates client-side — no request hits a server.  Last sync: 2 d ago.

Start a project
The architecture

Not a chatbot — an organization of agents.

SDEN.AI is structured like a company: an executive board sets direction, teams execute, and ~30 sub-agents specialize. A deterministic backbone routes the work; judgment lives at the nodes; humans hold the gates. 51 agents in all.

Foundations

Five principles

01

Clear hierarchy

An executive board sets direction, teams execute, sub-agents specialize.

02

Deterministic backbone, judgment at the nodes

The flow — who does what, in which order — is driven by the orchestrator and the kanban; cognitive decisions are delegated to the agents.

03

Skill-based routing

Every agent carries a description; the orchestrator routes each task to the best-placed agent.

04

Quality & governance

A critic agent verifies; human gates block any irreversible, external or costly action.

05

Shared memory

Every agent reads and writes a shared Obsidian vault (reports, wiki) with semantic search (RAG) on top.

Org chart

The layers of the organization

1Owner (human)Sets the objectives, approves sensitive actions, receives the reports.
7Executive boardTranslates the strategy, supervises the divisions — reports directly to the owner.
2OrchestrationThe orchestrator decomposes and routes; the critic guards quality.
11Team leadsDrive a domain and coordinate their own sub-agents.
~30Sub-agentsExecute a single specialty — analysis, writing, tests, publishing…
Roles · who does what

Executive board

ceo

Global strategy, priority arbitration, single point of report to the owner.

Synthesizes the board's briefs, decides, inserts the gates.

coo

Execution & process, delivery.

Drives the orchestrator + the kanban, tracks throughput and blockers.

cto

Technology — ops, coder, techdoc.

Technical health, roadmap, risk.

crdo

R&D and innovation.

Drives the rnd team: explore, prototype, evaluate.

cmo

Marketing — X, LinkedIn, Reddit, SEO, content.

Acquisition, channel performance, brand.

cro

Commercial.

Pipeline, ICP, conversion, revenue.

cfo

Finance & cost.

Token/infra budget, ROI, spend guardrails.

orchestrator

Decomposes the objective, routes, sequences, gates, tracks progress.

Kanban (decomposition + routing by description), delegation.

critic

Adversarial verification, quality gate.

Hunts inconsistencies and hallucinations; renders a verdict before “done”.

Platform

CTO
  • opslead

    Infra/DevOps lead — VPS health, security, backups. — Read first, plan + rollback → ops-monitor, ops-deploy.

  • ops-monitor

    Observability, incidents (read-only). — Prometheus / Grafana / Langfuse → alerts ops-deploy.

  • ops-deploy

    Deployments, releases, config-as-code. — Dry-run + rollback; prod action → human gate.

  • coderlead

    Dev lead — design, review, delivery. — GitHub (MCP) → coder-impl, coder-tests.

  • coder-impl

    Implementation (sandbox). — Small changes; → coder-tests.

  • coder-tests

    Tests / software QA. — Pass/fail verdict; blocks the merge on failure.

  • wikicuratorlead

    Knowledge lead — writes the wiki/vault. — HERMES.md conventions; leans on researcher.

  • researcher

    Research, mapping, ingestion. — Monitoring + RAG → wikicurator.

  • contentlead

    Content / Comms / Business lead. — Editorial plan, n8n → writer, comms.

  • writer

    Long-form writing. — → critic, then comms.

  • comms

    Multi-channel distribution. — n8n / Telegram; external send → gate.

Marketing

CMO
  • mkt-xlead

    X / Twitter lead. — Coordinates analysis → writing → publishing.

  • mkt-x-analyst

    Competitive monitoring on X. — Firecrawl (live web) → mkt-x-writer.

  • mkt-x-writer

    X posts & threads. — Hook + CTA → critic.

  • mkt-x-pub

    Publishing & engagement on X. — n8n; never publishes without a gate.

  • mkt-li + subs

    LinkedIn (analyst, writer, pub). — Same flow, B2B codes.

  • redditlead

    Reddit lead. — Coordinates monitor / analyst / writer / engage.

  • reddit-monitor

    Subreddit & mention monitoring. — Alerts with links and context.

  • reddit-analyst

    Trends, sentiment, competition. — → reddit-writer.

  • reddit-writer

    Native (non-promo) content. — → critic.

  • reddit-engage

    Community / publishing. — n8n; gate before any post.

Search (SEO / GEO)

CMO
  • seolead

    Global search lead. — Prioritizes, coordinates the specialists.

  • seo-analyst

    Keywords, intent, SERP, rankings. — Data → seo-content / technical.

  • seo-technical

    Audit (crawl, CWV, schema.org). — Fixes → ops/coder (gate).

  • seo-content

    On-page, internal linking, topical authority. — Briefs → critic.

  • seo-offpage

    Link building / authority. — Opportunities; external contact → gate.

  • geo-specialist

    Optimization for AI engines. — Citable content (ChatGPT / Perplexity / AI Overviews).

Sales, R&D & Documentation

CRO · CRDO · CTO
  • saleslead

    Commercial lead (GTM, ICP, pipeline). — Coordinates research / copy / crm.

  • sales-research

    ICP, prospects, competitive monitoring. — → sales-copy.

  • sales-copy

    Cold emails, sequences, DMs. — → critic → gate before sending.

  • sales-crm

    Pipeline, follow-ups, statuses. — n8n; no external send without a gate.

  • rndlead

    R&D lead. — Coordinates research / proto / eval.

  • rnd-research

    State of the art, sources. — → rnd-proto.

Mechanism

Lifecycle of an objective

Objective

set by the owner

orchestrator

decomposes + routes

sub-agents

execute, write to the vault

critic

verifies

Human gate

if action on the world

Report

into Obsidian

CEO

synthesis

Owner

decides next

How it works

The mechanisms

Decomposition

The orchestrator (or the kanban decomposer) breaks an objective into sub-tasks with dependencies.

Routing

Each task is assigned by description to the best-placed profile; leads delegate to their sub-agents.

Autonomous execution

The dispatcher built into the gateway (30 s tick) launches ready tasks — without intervention.

Memory

Everyone writes their deliverables to the Obsidian vault (reports/<team>/); search_vault (RAG/Qdrant) gives a shared semantic memory.

Quality

The critic verifies every non-trivial deliverable before “done” — anti-hallucination, format, feasibility.

Human gates

Any irreversible, external, costly or prod action is parked for approval.

Model tiering

“Brain” roles run on claude-sonnet-4.6; executors on gpt-5.4-mini — cost kept in check by the CFO.

Sandboxing

Agent code runs in ephemeral containers, isolated from the secrets.

Governance & security

What keeps it safe

Gates by risk level

Green (read / analyze / draft) is autonomous; orange-red (prod, external send, spend, deletion) needs human approval.

Safe harness mode (A+B)

Autonomous workers only launch knowledge roles, with a restricted toolset; acting roles stay gated.

Anti-hallucination

Agents refuse to invent — an analyst blocks if it has no source rather than hallucinate.

Observability

Langfuse traces tokens / cost / latency per agent; the CFO and a cost guard watch the spend.