I
MIND
04
Hyper-systematization
Hyperfocus
Explicit communication
Solo work
How I think: my thinking already has the shape of a spec
Roman Nekrasov
Personal Manifesto of an AI Researcher/Architect-Researcher in the Agentic Era
When agents write the code, value moves from the ability to write to the ability to frame the problem, hold the context and prove the result on real hardware. These seven advantages are my answer to that shift.
THE MOAT · three layers, each amplifying the next
A single tool can be picked up in a week. The product of mind × craft × proving ground cannot be bought — only accumulated, run after run.
II
CRAFT
03
Terminal harnesses
Browser harnesses
Context library
What I command: the agent as a runtime, context as capital
III
PROVING GROUND
07
GB10
Ubuntu
Android
vast.ai
Opus 5.5
Kimi K3.1
Fable 5.1
Where I prove it: live runs and three independent minds
A single tool can be picked up in a week. The product mind × craft × proving ground cannot be bought — only accumulated, run after run.
Four pillars
- why I chose it
- what it unlocks
- self-realization
01
Cognitive style
WHY I CHOSE IT
An agent is a literal executor: it does not infer the unsaid or read between the lines. A mind that structures the world and speaks plainly matches the agent's interface one to one.
WHAT IT UNLOCKS
My first prompt is already a specification. Hyperfocus carries a multi-hour session without losing the thread. Working solo removes coordination overhead: decision, run, conclusion — one loop.
SELF-REALIZATION
Traits once seen as "inconvenient" become a professional edge. Not bending to the environment — finding one built the way I think.
02
Terminal harnesses & browser
WHY I CHOSE IT
The model is the engine; the harness is the chassis. Results come from the whole rig: tools, permissions, verification loops, subagents. The terminal executes; the browser reaches the real web and live UI.
WHAT IT UNLOCKS
I drive the agent as a runtime, not a chat: I know where a harness breaks, how to split a task, when to fork parallel branches and how to close a self-check loop.
SELF-REALIZATION
From "I write code" to "I design the production system that writes code". That is architecture in the agentic era.
03
Context library
WHY I CHOSE IT
Context is the new codebase. Skills, templates, house styles, reference examples and proven prompts form an asset that grows and is reused in every next run.
WHAT IT UNLOCKS
Every run starts from accumulated capital instead of an empty window. Compound interest: each project makes the next one faster and sharper.
SELF-REALIZATION
Knowledge no longer vanishes with the context window. It becomes a legacy that keeps working even when I am not in the session.
04
Live runs on my own proving ground
GB10local inference, privacy
Ubuntuserver reality
Androidmobile, on-device
vast.aiGPUs on demand
WHY I CHOSE IT
For a researcher, an opinion without a run is a hypothesis. Four platforms span the whole range — from private local hardware to an elastic cloud.
WHAT IT UNLOCKS
I bring logs, metrics and a reproducible run. In a large organization that turns "it seems" into "it is measured".
SELF-REALIZATION
Seeing the system behave on real hardware with my own eyes, not retelling someone else's benchmarks.
Three minds, not one
No single model holds a monopoly on truth. Three independent schools let me triangulate answers, route tasks by cost and quality, and stay free of any one vendor's blind spots.
Opus 5.5
ROLE · DEPTHHard reasoning, architecture decisions, long agentic sessions inside the harness.
Kimi K3.1
ROLE · SECOND OPINIONA different training school: cross-checking, harness comparisons, cost control.
Fable 5.1
ROLE · FRONTIERMythos-tier with added safeguards: the ceiling of capability and an early look at what comes next.
THE FLYWHEEL · why the sum beats the parts
- 01 Mind a precise frame
- 02 Harness a fleet executes
- 03 Three minds cross-check
- 04 Proving ground proves on hardware
- 05 Library remembers ↺
OPERATING PRINCIPLES
/01
Say it explicitly, or the agent won't know it.
/02
No run, no claim.
/03
Every lesson goes into the library.
/04
Two models agree before I trust one.
THE PATH
NOW
Source of proof
The person to ask when an answer must be tested on real hardware.
NEXT
Author of agentic practice
The context library becomes a standard for the organization.
GOAL
A system larger than one person
An architect whose thinking scales through a fleet of agents.
RISK
Solo risk — isolation. Antidote: publish runs and findings.
RISK
Focus risk — tunnel vision. Antidote: a ritual of reviewing goals.
RISK
Vendor risk — lock-in. Antidote: three models and my own library.