
Fernando Ramos
Co-Founder, Co-CEO & CFOAgentic Engineering
Fig. 1. One orchestrator, specialised agents, their subagents. Every edge is observable.
Selected work
- 01CuriaLegal AI platform2025–nowAI for Mexican lawyers: cases, sources and human review in one workspace.
- 02Production AIEnterprise automation2024–2025Voice and WhatsApp automation in daily operations, delivered with a team of 12 developers.
- 03SensorGridResearch manuscript2026Monitoring physical systems so trouble shows up as a pattern before it becomes a breakdown.
- 04MemSWEBenchmark research2026A benchmark that tests agent memory on real software maintenance, with everything else held constant.
Method
Agents are not chat windows.
One orchestrator manages specialised agents, and each of them manages its own subagents. There is no best model and no best harness, so I use each for what it does well, and I only scale a way of working once it is standard and trusted. Six rules keep it honest.

- 01
Observable
An agent you cannot see is an agent you cannot improve, correct or trust.
- 02
Not a clipboard
Do not be the person who copies one agent’s output into another. Let them talk to each other.
- 03
Evidence of done
Saying it is finished is not proof. Ask to see the work, and that it is right.
- 04
Context continues
Context windows are finite. Design so the work survives them, and use them in your favour.
- 05
Many models, many harnesses
There is no best model or harness. Give each the work it is good at, like a team with different strengths.
- 06
Autonomy with ownership
Agents get bounded autonomy and come to you when blocked. You stay responsible for what they do.
Open source
Tools for working with agents, in public.
SpecSafe
MITSpec-driven workflow
A skills-first framework that keeps coding agents on the intended outcome: spec, test, code, QA, complete. Works across Claude Code, Cursor, OpenCode, Aider, Zed, Gemini and more.
Odin
MITAgent command center
An overlay on the Pi coding agent that adds fleet management, handoffs and multi-agent orchestration, while leaving upstream Pi untouched.
Seshat
Public repositoryAgent harness
A custom Pi harness with persistent memory, the SpecSafe slice lifecycle, per-slice cost accounting and agent-aware git traceability.
