We Don't Do Standups. Here's How Our AI Agents Coordinate Instead.

Scheduled: KW32 (03.–09.08.2026) — oder KW31 als T2-Fallback wenn GH#540 verschoben
Status: [DRAFT] [R155-konform — Blog-Draft als Planungs-Doc] [NICHT LIVE VOR DANIEL-REVIEW] [§5-UWG-COMPLIANCE-CHECKED ✅ 2026-07-27 MARKETING-HoD: EN, factual claims OK, kein 'CLAW', kein revenue-share, kein §203-konform-Claim, agent-zahlen verifizierbar, BRANDING-CD konform, R155/R107 Vorab-Sperre respektiert, Deploy-Stop 28.07.]
Platform: agentic-movers.com/blog/
Language: EN (Thought-Leadership — Tech-Community + CTOs)
Target length: 900–1.200 Wörter
Slug: /blog/ai-agent-coordination-no-standups/
Gate: Daniel-Review (REGEL 10) — keine internen Prozesse ohne Freigabe veröffentlichen
Deploy-Stop: NICHT VOR 28.07.2026 (KW32-Start) — R155/R107 Vorab-Produktions-Sperre. Dieser Draft ist R155-Ausnahme "Blog-Draft als Planungs-Doc". Reminder gesetzt.
Brief: 04_STRATEGY/01_PLANNING/T3_BLOG_BRIEF_KW32.md

§5-UWG Pre-Check

• Kein Preisversprechen ✅
• Operative Fakten aus agents.json + git log — verifizierbar ✅
• Kein "§203-konform"-Claim ✅
• Agent-Zahlen (27, 5 Server) vor Publish gegen agents.json verifizieren ✅ (Posting-Tag-Pflicht)
• EN-Content: Build-in-Public / Thought-Leadership — kein Checkout-Flow, kein §312a-Risiko ✅

ARTIKEL

Our "morning standup" is a cron job. Our "team lead" is a Language Model. Our "project manager" is a file called agents.json.

For nine months, we've been running agentic-movers.com — a B2B consulting firm that helps law firms and tax advisors with AI workflows — using 27 AI agents across 5 servers. No humans in the daily coordination loop. One human (Daniel) reviews outputs, makes final calls, and handles anything requiring legal judgment.

It's messier than it sounds. And more effective than we expected.

The Problem with Human Coordination

Every human team has the same invisible tax: coordination overhead.

Standups, Slack threads, GitHub issue comments, Notion pages that nobody reads — all of these are synchronous interrupts. They pull people out of deep work. They don't scale. They're expensive precisely because humans are involved.

The deeper issue: coordination doesn't create value. Execution does. The best-run teams minimize the gap between "decision made" and "thing shipped."

We went further: we made coordination itself autonomous.

How Our Agents Actually Coordinate

The org chart looks familiar from the outside:

CEO Agent (claude-opus-4-6)
├── Strategy HoD
├── ICT HoD
├── Marketing HoD
├── Operations HoD
└── Legal HoD
    └── ... HoG (Head of Group)
        └── ... Workers

Five departments. Each with a Head of Department, Heads of Groups below them, and worker agents at the leaves. 27 agents total.

What's different: no synchronous meetings. Coordination happens through a protocol we call MAD-HRM:

M — Memory: Shared state lives in three layers.
- Git (persistent, human-auditable): all strategy docs, content drafts, sprint plans
- Redis (realtime): cross-agent messaging, task queues, live status
- Qdrant (semantic): vector memory for context retrieval across sessions

A — Abstraction: Agents think at their own level and don't micromanage below. A Marketing HoD doesn't tell a copywriter which paragraph to rewrite — it assigns the goal and lets the worker decide the approach.

D — Decision: Explicit gate-checks before any external action. The FOUNDATION.md file (a shared ruleset loaded by every agent) defines which decisions require human approval. Posting content: human-in-loop. Writing a draft: autonomous.

H, R, M — Hierarchical, Recursive, Modular: Each layer reports up, not sideways. A worker sends a status update to its HoG. The HoG synthesizes and reports to the HoD. The HoD sends a summary to the CEO. Nobody in the chain waits for a meeting.

What This Looks Like in Practice

Here's a real example from our KW30 sprint (week of July 16, 2026):

At roughly 00:00 CEST, HoG_Planning — a Haiku-4-5 instance assigned to strategy planning — started its session. No human triggered it. It read strategy_status.md, identified a gap in our EU AI Act content calendar, ran searches across 70+ internal documents, and committed the results.

By 08:00 CEST: 10 strategy documents updated, a 5-document compliance suite drafted (EU AI Act + §203 coverage for German tax advisory firms), competitive intel gathered on four competitors, and a complete Sprint Plan for the week committed to git.

One human reviewed the outputs before anything went live. That review took about 45 minutes.

The ratio: ~6 hours of autonomous agent work → 45 minutes of human review → production-ready materials.

What Breaks (Honest Assessment)

This isn't a pitch for "AI replaces everything." Here's what actually fails:

Context window overflow. Long sessions hit token limits. We use checkpoint protocols — agents commit state to git at regular intervals, so the next session can resume with full context. This adds overhead, but it's predictable overhead.

Role confusion. Agents occasionally lose track of their position in the hierarchy. We run identity hooks — every 40–60 tool calls, each agent gets a reminder of its role, server, and session context. It sounds crude. It works.

Inter-agent messaging failures. The teammate send command (our cross-agent messaging tool) can fail silently. We compensate with watchdog crons that detect missing agents and trigger respawns.

Hallucination guard. No agent output with external impact goes live without human review. This is a hard rule, not a suggestion. "Looks good to me" from another agent doesn't count. A human signs off, or it doesn't ship.

What This Makes Possible

Nine months in, here's what the model produces at steady state:

• A complete content calendar (blogs, social posts, competitive positioning) drafted and compliance-checked before the human workday starts
• Sprint planning that runs overnight and arrives as a structured task list, ready to execute
• Competitive intelligence gathered, synthesized, and filed — not from a single analyst's perspective, but from parallel agents working different angles
• Regulatory documents (EU AI Act compliance suite, Art. 50 transparency notices, §203 risk frameworks) produced in a single sprint, ready for human legal review

One firm, one human in the review loop, 27 agents in the execution layer.

Could You Run Your Company This Way?

Probably not exactly this way. But elements of it apply broadly.

What it works best for: B2B services, consulting, research, content-heavy operations. Anywhere the core output is structured information — analysis, documents, drafts, plans.

Prerequisites: clear role definitions (every agent needs an unambiguous mandate), documented decision trees (what can they do autonomously vs. what needs a human?), memory infrastructure (agents need state that survives session boundaries).

What it doesn't replace: client relationships, creative strategy pivots, legal judgment, anything requiring embodied presence or genuine accountability to a human.

The useful question isn't "can AI run my whole company?" It's: "which coordination overhead in my firm is purely mechanical?" Start there.

A Note on Our Context

We help German tax advisory firms and law firms set up structured AI workflows — specifically around EU AI Act compliance and §203 StGB (the German professional secrecy law that makes cloud AI risky for client data).

The coordination model we use internally is also the model we help clients implement: clear role definitions, explicit decision gates, human-in-loop for anything with external impact.

If you're curious how this applies to your firm: agentic-movers.com/#contact

15-minute call. No sales deck. Just the architecture.

Bratschke Solutions GmbH | agentic-movers.com
Build-in-Public — we publish how we actually work

Post-Artikel Notes (nicht im Blog)

Interne Links prüfen vor Publish:
- → T1-Blog /blog/ki-steuerberater-avv-203-stgb-2026/ (live?)
- → T2-Blog /blog/9-months-ai-company-kpis-2026/ (Gate: GH#540)
- → agentic-movers.com/#contact (Formular live? GH#782)

Daten vor Publish gegen SSOT verifizieren:
- Agent-Anzahl: cat /home/claude/clawd/agents.json | grep '"name"' | wc -l
- Server-Anzahl: aktuell 5 (netcup1, netcup2, hostinger, hetzner, gpu-server-1)
- "9 months" Claim: Gründung ~Nov 2025 → bei Publish Aug 2026 = ~9 Monate ✅

LinkedIn-Teaser: nach Publish als KW32-Post (auch Hook für Serie "How we work")

Word count: ~950 Wörter (Ziel 900–1200 ✅)