What Happens When You Brief 22 AI Agents on a Single Task

Matthias Heim · 2026-03-13

1,100 sessions, 430M tokens, 22 full-time knowledge workers. Cost: CHF 200/month. Most of it was waste, and that's the point.

Last month: 1,100 agent sessions. 430 million tokens processed. The equivalent of 22 full-time knowledge workers. Total cost: CHF 200, or CHF 9 per "employee." Most of what they produced was waste. And that's exactly the strategy.

The Setup

This didn't start as a grand plan. It grew organically from a single email drafting agent. One agent became three, three became ten, and now there are 22 running daily.

Mix of launchd schedules, calendar sync triggers, and Claude Code sessions

Markdown files, Python scripts, Neon Postgres for shared state

No LangChain, no CrewAI, no agent frameworks: just scripts and prompts

Each agent has a single responsibility and a markdown spec file

What 22 Agents Actually Do

Before I Wake Up

Memory consolidation: reviews yesterday's work, updates project context files

Security audit: scans repos, checks dependencies, reviews access logs

LinkedIn content: drafts posts from my recent work and bookmarks

Email triage: categorizes inbox, drafts replies for routine messages

Throughout the Day

Calendar sync: prepares briefings for upcoming meetings

Meeting prep: researches attendees, pulls relevant context from past meetings

Transcript processing: syncs and analyzes completed meeting recordings

Email monitoring: drafts responses as new messages arrive

On-Demand

Lead research: deep-dives on prospects before outreach

Content creation: blog drafts, newsletter sections, case studies

Client-specific agents: custom workflows for individual client needs

The Waste Problem

Let's be honest about the numbers:

Security audit ran 28 times last month. Found 1 actual issue.

Meeting prep for 34 meetings. 6 were cancelled. Only about 60% of the prep was actually useful.

Lead research on 47 prospects. 8 responded. Rest was wasted research.

Email drafts: 25 out of 40 were used. 15 were rewritten or discarded.

Content: more thrown away than kept. Maybe 30% survives to publication.

Overall: 30-40% of agent output is directly useful. The rest is waste.

Why Waste Is the Strategy

Here's the key insight: at CHF 9 per agent per month, the economics are completely different from human employees.

That one security issue the audit found? It was a misconfigured access token that could have been catastrophic. Worth 27 empty reports.

Research all 47 leads instead of guessing which 8 will respond. The cost of researching 39 dead ends is basically zero.

Prep all 34 meetings because cancellations are unpredictable. Being over-prepared costs nothing.

Audit every day because threats don't announce their schedule.

When marginal cost approaches zero, the strategy shifts: think about everything, filter after. The waste IS the moat.

What Actually Failed

Not everything is rosy. Here are genuine failures:

Context Drift

Agents working from stale assumptions. One agent kept referencing a pricing model we'd abandoned weeks ago.

Conflicting Outputs

Two agents researching the same prospect produced contradictory recommendations. No resolution mechanism.

Quality Ceiling

Creative work hits a wall. LinkedIn posts and blog intros are competent but lack a distinctive voice without heavy editing.

Debugging Cascades

When one agent's output feeds another and something breaks, tracing the issue across 4 agents is painful.

Scheduling Collisions

Multiple agents hitting APIs simultaneously. Rate limits, context window conflicts, resource contention.

Cost Breakdown

~CHF 150

~CHF 15

~CHF 20

~CHF 10

~CHF 5

~CHF 200

CHF 9 per agent/month

The Real Bottleneck: Me

The agents aren't the limiting factor. I am.

Coordination overhead: 1-2 hours per day reviewing outputs, fixing issues, updating specs

Every new agent needs a markdown spec, test runs, and monitoring setup

Can't scale to 50 agents without better tooling for orchestration and conflict resolution

The dashboard I should have built on day one still doesn't exist

Advice If You're Starting

Start with one daily agent

Email triage or meeting prep. Get the rhythm right before scaling.

Invest in shared context

A single Postgres database that all agents read from and write to. Without shared state, agents are islands.

Accept waste early

Don't optimize for precision on day one. Let agents be thorough and filter the output yourself.

Build a dashboard early

I wish I had. Knowing which agents ran, what they produced, and what failed saves hours of debugging.

Keep humans in the loop for outbound

Let agents draft, research, and prepare. But anything that goes to a client or prospect gets human review.

How we work · Work · Insights