Manuel Porras
38 agents in operation

I build AI agents that run in production, not demos.

I decide which processes are worth handing to an agent, give each agent the knowledge, tools and clear limits it needs, connect it to the company’s systems — and make sure it runs around the clock and recovers on its own.

Manuel Porras
Manuel PorrasAI Agent Engineer · Berlin · Available for projects and permanent roles
See the workCV
⌘ K
Example questions · fictional client “Acme HVAC”
fleet · 38 agentsNavigate ↑↓Open ↵
38Claude Code agents in 24/7 operation
18client-expert agents, one per client
1 minwatchdog interval — crashed agents rebuild automatically
2022working with generative AI every day since
Measured on the production servers, October 2026.

Three systems in daily use

01 / production

Built for a US marketing agency serving home-service businesses. Clients stay anonymous; the numbers are real.

Agent fleet · Client Control
Case study · Agent operations

One Slack channel, 18 expert agents behind it

An agency team asks questions about its clients in Slack. A fleet of agents answers — each one an expert on a single client.

  • scale38 Claude Code agents on two servers, one expert agent per client.
  • routingA router agent sends each question to the right client agent.
  • knowsMeetings, tasks, conversations and four ad platforms of its client.
  • safetyRead-only for the team — enforced in code, not in the prompt.
  • uptimeA watchdog checks every minute and rebuilds crashed agents.
Read the case study
Meeting intelligence
Case study · Human in the loop

From client meeting to approved task in ~15 minutes

Account managers used to type up tasks by hand after each call — items got lost, with delays of up to 2 days.

  • triggerThe finished Zoom transcript starts the pipeline; polling as a safety net.
  • proposeAn LLM proposes tasks with likely owners — also for client requests in Slack.
  • approveA human approves every proposal before anything is created.
  • sinceIn daily use since April 2026, for 30+ client meetings a week.
Read the case study
PPC monitoring console
Case study · Daily operations

One look every morning instead of a login per ad account

The whole client portfolio and each client: budgets, cost per lead, lead sources and tracking — with trends, traffic lights and alerts.

  • dailyMy first look every morning — weekly and monthly trends, not just today.
  • alertsCost per lead too high or too low, budget pacing, a card that needs renewing.
  • safetyRead-only on every ad platform; aggregates only, no personal data.
  • honestMissing data stays missing; a failed run never deploys.
  • auditReal case: the real cost per lead was about 60 % above what Google reported.
Read the case study

All projects

02 / projects
Projects
Agent fleet · Client Control One Slack channel, 18 client-expert agents behind it, read-only by designagents · 2026 Meeting intelligence Client meetings become approved tasks in about 15 minutes, 30+ meetings a weekagents · 2026 PPC monitoring console Every morning: budgets, cost per lead and tracking for all clients in one view.infrastructure Search-visibility audit Where is a local business invisible on Google? Evidence per company, under $0.10 each.agents · 2026 Address → 3D scene Agents write Blender code from geodata and photos — 30+ 3D sets in 10 days.creative lab · 2026 Self-healing agent fleet 38 AI agents running unattended on two machines, recovering on their ownagents · 2026 Private call analysis 266 calls transcribed on my own GPU — and an assumption disprovedinfrastructure · 2026 Blind evaluation harness Blind-test harness for AI output: comprehension 3.3 → 7.8 over 56 cyclesevaluation · 2026 White-label AI platform 2025: 30+ AI assistants in one environment, new clients by configuration, not codeinfrastructure · 2025 Ad anomaly detection Earlier project: daily anomaly alerts for the ad accounts of 30+ agency clientsinfrastructure Ask the agent fleetRouter, read-only tools and a refused write attempt — test data.live demo Meeting → tasksApprove, reject, merge: the review step of the meeting pipeline — test data.live demo

How I work

03 / principles
Settings · Principles
01
A human approves where a mistake costs trust

Agents propose, people confirm. Each decision becomes data, so clear cases can be automated later.

02
Agents may only do what they need to

Permissions are set in code, per client — read-only by default.

03
Every change is measured

Blind tests instead of gut feeling: the judge never sees the source. 56 documented test cycles.

04
Systems have to survive the night

Watchdog, automatic rebuilds, restart-loop protection, and knowledge kept in files — nothing is lost after a restart.

More about how I work

Creative AI Lab

04 / design with agents

Agents do not only answer questions. Given the right tools, they build things you can look at.

cafe_neukoelln.glb
Experiment · Berlin in 3D

Berlin in 3D

From two photos of a café in Neukölln, my agents built an explorable 3D scene — by writing Blender code, rendered locally on my own GPU.

Open the 3D scene
Case study

From an address to a 3D scene

  1. Address
  2. Nominatim
  3. OSM · Overture · LoD2
  4. Blender, headless
  5. GLB
  6. Viewer

Streets come from open geodata, interiors like the café from photos — always as Blender code. AI mesh generators were tested and rejected.

Read the case study