Coming Soon Open Source MIT Licensed Free Forever

Seed Agent

Your digital cortex, powered by Claude Opus.

DIGITAL CORTEX. EVOLVING. LEARNING. SCALING.

A minimal, domain-agnostic foundation that turns Claude Code into a disciplined, self-improving assistant. You customize it through conversation — it adapts its own architecture to fit how you work.

New to seed agents? Start Here explains what they are and how they fit into your work.

View on GitHub (public soon) Back the Project
A profile portrait with a glowing brain visible inside the head, surrounded by luminous nodes labeled OBSERVE, PLAN, EXECUTE, VERIFY, CONDENSE, and IDLE — visualizing the OPEVC cognitive cycle.

What It Is, in Plain Words

The Seed Agent is a small set of files and scripts you drop into a folder. Once installed, it changes how the LLM behaves: instead of jumping straight to action, the agent observes, plans, executes, verifies, and learns — every time. It remembers what you taught it, organizes its work as jobs, and refuses to stop in the middle of a task.

It is not a chatbot, not a SaaS product, and not behind a paywall. It is a starting template — a seed — that you grow into your own personal assistant by talking with it. A lawyer's seed agent and a researcher's seed agent start identical and end up looking nothing alike.

The seed agent ships with mechanisms, not categories. You bring the domain expertise; the agent provides the cognitive structure.

The OPEVC Cognitive Cycle

Every job the agent works on flows through five compartmentalized phases. Each phase restricts the agent's tool access, forcing it to think before it acts and verify before it finishes.

O

Observe

Gather context. Read files. Search the web. Pull what's needed into working memory.

Outputs: synthesis notes
P

Plan

Design the approach. Define strict criteria. Decide what success looks like before any code runs.

Outputs: implementation plan
E

Execute

Implement the plan surgically. No improvising. The plan is the contract; execution fulfills it.

Outputs: file changes + logs
V

Verify

Run validation. Check every acceptance criterion. No "looks good" judgments — only pass or fail.

Outputs: pass/fail verdicts
C

Condense

Distill what was learned. Move durable knowledge into long-term storage. Reset working memory.

Outputs: knowledge files

This rhythm prevents the agent from skipping vital deliberation, hallucinating shortcuts, or losing what it learned the moment a session ends.

Jobs Mature Over Time

Everything the agent does is tracked as a job. Jobs come in three formats, called Stages — each a valid starting point. As the agent gets better at a kind of work, you and the agent decide together when the next run deserves a more structured Stage.

Stage 1

Single-Cycle

highly collaborative

For brand new tasks. The agent runs through one OPEVC cycle, going back and forth with you to figure things out. At the end, a single Condense phase absorbs everything learned. This is how the agent learns a job for the first time.

Stage 2

Multi-Cycle

structured practice

For repeatable, structured work — often once a pattern is understood. The agent writes a dedicated plan file, breaks the work into chunks, and runs multiple smaller OPEVC cycles against it. The plan persists and gets smarter with every run. The job becomes repeatable.

Stage 3

Multi-Cycle, Self-Briefing

matured practice

Identical to Stage 2 in how it runs and completes — but its plan injects job-specific guidance into every working phase, tailoring the agent's coaching to this exact work. The job briefs the agent at every step. (On the roadmap: a fourth stage where a fully understood job ships its own plugin.)

What Ships in the Seed

Eleven hardened plugins form the agent's identity at install time. They handle phase enforcement, job lifecycle, behavioral discipline, and structural integrity — all the cognitive scaffolding so you can focus on the work.

plugin_integrity

Single-plugin editing, test-verified locking — the foundational protection layer every other plugin depends on.

job_core

Job lifecycle, stop-blocking, dependency management. Refuses to let the agent quit mid-task.

phasic_system

Idle enforcement. Phase and cycle tracking. The orchestrator behind every OPEVC transition.

phase_observe

Guards the Observe phase — read-only access, depth gates, structured note-taking.

phase_plan

Guards the Plan phase — multipliers, plan files, acceptance criteria enforcement.

phase_execute

Guards the Execute phase — write access only to scoped directories, plan adherence checks.

phase_verify

Guards the Verify phase — runs acceptance scripts, blocks "looks good" judgments.

phase_condense

Guards the Condense phase — runs the 7-step waterfall that turns experience into long-term knowledge.

interaction_summary

Captures every back-and-forth so user intent never gets lost between phases.

brain_guard

Behavioral enforcement. Catches drift, prevents context bloat, keeps the agent disciplined.

question_discipline

Structured questioning — one question at a time, well-formatted, multiple-choice when possible.

How You'll Install It

When the public repo goes live, installation is three steps. Designed for fresh, empty projects so nothing of yours gets overwritten.

1

Clone the repo

git clone https://github.com/hadi-nayebi/seed_agent.git
cd seed_agent
2

Run the installer

./install.sh /path/to/your/new_project

Copies the digital cortex into a fresh project. Refuses to overwrite an existing one.

3

Talk to it

cd /path/to/your/new_project
claude --allow-dangerously-skip-permissions

Then say what you want: "I want to build a landing page for my bakery." The job system takes over.

Roadmap

The seed agent is engine-agnostic by design. Today it runs on Claude Code. After the public release ships, ports to OpenCode and Gemini follow.

Today

Claude Code

The first home of the seed agent. Powered by Claude Opus. Final preparations before the public repo lights up.

Next

OpenCode

Same OPEVC cycle, same job system, ported to OpenCode's tool surface. Begins after the Claude Code release.

Then

Gemini

The third engine. Same cognitive scaffolding, adapted to Gemini's runtime. Same agent, three brains.

A full Roadmap page is in the works. Until then, this is the order of what's coming.

Ready When You Are

The public repo lights up shortly. Until then, you can star the placeholder, join the community, or back the work.

Watch the Repo

Star or watch the GitHub repo to know the moment the public version drops.

Open the repo →

Join the Community

Talk with other professionals already exploring agent engineering. Questions, patterns, feedback.

Join on Skool →

Back the Project

The seed agent stays free forever. If you want to support the work that keeps it that way, it helps.

Support the project →

This project is free and self-funded. If it helps you, you can back the project — entirely optional.