Skip to main content

Lab Notebook · Everything-as-agent

Everything is an agent — and the agents write their own tools

Through 2025 we built a platform where every component is an agent with a standard contract, and a system (SelfGrow) where agents create the skills they lack instead of failing. The agent contract became the foundation of everything that followed. Honest edge: self-growth works for small increments and still needs a human for the big architectural calls.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 20 May 2025
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 20 May 2025
Living document. Reviewed 20 May 2025
Entry date
20 May 2025
Category
Agents
Lead over the world
~1–2 years ahead
Access
🔓 Public
ResearchPartly provenDownload PDFPDF · 737 KB

An agent that hits a task it can't do usually fails. Ours wrote the skill it was missing and carried on — and every part of the platform, down to the gateway, was itself an agent.

v1→v3
SelfGrow iterations, CLI to live dashboard
5-part
standard agent contract (blueprint/code/settings/types/GUI)
13 in 48h
network architecture versions, AI-paced

Honest evaluation

Partly proven

The five-part agent contract fully proved out and became foundational; self-growth (agents writing their own missing skills) proved real but bounded — reliable for small increments, not for architectural pivots.

What would prove or disprove it further

What would extend the self-growth claim toward fully proven: an agent-generated skill that triggers, unprompted, a restructuring of its own architecture that a human later validates as correct — the thing the report says didn't happen. Short of that, "self-growth works for increments" is as far as the evidence goes.

The evidence — full reasoning behind the verdict

Verdict: partly-proven. Two separate claims are bundled in this report, and they landed differently. The five-part agent contract (blueprint, code, settings, types, GUI) is the clean proven half: it made agents composable and swappable, and the report states plainly that it "survived everything and became foundational" — later work, including the production Director → Manager → Worker → Tool hierarchy, is built on it. The self-growth claim — agents that write their own missing skills instead of failing — is the partial half: it worked, but only within a bounded scope.

The mechanism is stated honestly rather than glossed over: self-growth is reliable for incremental skill gaps, where the agent needs one new tool and writes it. It has no answer for architectural pivots — deciding the system should be restructured is still a human call, not something SelfGrow ever did on its own. The thirteen-versions-in-48-hours pace is real evidence of fast, AI-assisted exploration, but the report is equally honest that this speed produced maintenance debris rather than a free lunch.

The conceptual breakthrough of 2025 was small to state and large to build: make every component an agent, give agents a standard shape, and let them write the tools they don't have. Most of what vocabotics is grew out of that sentence.

What it was

Two ideas, developed together:

  1. The agent contract. Instead of ad-hoc scripts, every agent had the same five-part structure — a blueprint.json, the code, settings.json, types.ts, and a gui.tsx. Standardising the shape made agents composable and replaceable: you could swap one for another, generate new ones on demand, and reason about the whole system uniformly.
  2. Self-growth. An agent that meets a task outside its abilities normally fails. SelfGrow instead had the agent create the missing skill on the fly and continue — learning by extending itself rather than stopping.

What we built

  • SelfGrow v1 → v3: three iterations, from a CLI to a real-time Socket.IO web dashboard that visualised skills as the agent acquired them.
  • vocabotics Agents: the platform that made the five-part contract concrete, with a CLI, tab-completion, and dynamic agent generation.
  • vocabotics Network v1 → v13: thirteen architectural versions in two days, each testing a different boundary — agent registry, identity, networking, even an LLM-based proof-of-work experiment.
  • Aigen / Grow: pushed "everything is an agent" to its logical extreme — the API gateway and the marketplace were also agents.
Research

The measured signal here is less a single number than a rate: thirteen coherent architectural iterations in forty-eight hours is a pace that is simply not available without AI-assisted development, and it let us explore the agent design space empirically instead of on paper.

What we learned — including the honest negatives

  • Self-growth has a ceiling. It works well for incremental skills — the agent needs a small new tool and writes it. It struggles with architectural pivots: deciding to restructure the whole system is still a human judgement call. An agent can grow a limb; it can't yet decide it should have been a different animal.
  • "Everything is an agent" is not free. Making the gateway an agent, and the marketplace an agent, is elegant — and sometimes it is just indirection where a plain function would have been clearer and faster. The abstraction earns its keep in some places and taxes you in others, and knowing which is the skill.
  • Fast iteration leaves debris. Thirteen versions in two days is a superpower for exploration and a liability for maintenance — you end up with near-duplicates that later have to be pruned.

Where it went / status

The agent contract survived everything and became foundational. The self-growth idea matured into the lab's later work on a governed fleet, and the Director → Manager → Worker → Tool hierarchy (a separate, protected report) is the production form of it. This was the year "everything is an agent" stopped being a slogan and started being an architecture.

What is still open — kept visible

The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.

  • Self-growth is reliable for incremental skills; it struggles with architectural pivots, which still need human judgement.
  • 'Everything is an agent' adds indirection where a plain function would do — the abstraction helps in some places and taxes you in others.
  • Rapid iteration (13 versions in 48 hours) explores the space fast but leaves a trail of near-duplicates that later have to be culled.

Where this connects

Sources

  1. vocabotics project audit — SelfGrow v1–v3, vocabotics Agents, Network v1–v13, Grow (2025)vocabotics internal project history · as of May–June 2025

    We use cookies.