A repo-native agent framework for Claude Code

Agent
Framework.

Build AI colleagues that bring judgment—not just wait for prompts.

curl -fsSL https://agent-framework.sh/init.sh | sh

Run inside your repository. Then use /create-agent in Claude Code.

Context loaded Ready to declare priorities

Explore the system

Most assistants wait.

These agents load context, declare priorities, notice drift, and save what changed.

01 / THE SYSTEM

The agent is inspectable.

No hosted persona. No separate database. Identity, authority, momentum, and memory live in Markdown alongside the work.

The repository is the system of record.

agents/cato/tracked by git

role.mdpurpose, outcomes, boundariesidentity

soul.mdvoice, temperament, valuesidentity

autonomy.mdauthority by action typejudgment

actions.mdstanding work and ownershipmomentum

MEMORY.mddurable knowledge indexcontinuity

memory/standing + session contextcontinuity

playbooks/repeatable operating proceduresmomentum

STARTUP SEQUENCE

01Load role and soul
02Calibrate authority
03Reconcile actions
04Recall standing memory
05Declare today’s priorities

02 / CALIBRATED AUTONOMY

Trust is not a toggle.

Authority is set per action type, then promoted or demoted through explicit experience. The goal is the right level—not maximum independence.

Flag

“I see a problem…”

Recommend

“I recommend…”

Intend

“I intend to…”

Act & inform

“I’ve done…”

Own

“Reported at close”

03 / EVERY SESSION

LoadPrioritiseWorkCompass checkSave

The agent drives. You steer.

04 / WORK WITH THE SYSTEM

Four commands. Durable context.

01/create-agent

Design a role, scope, autonomy, and personality before files are written.

02/agent <name>

Activate one colleague with its saved context, actions, and recent memory.

03/agent status

Build a live status board directly from every agent’s tracker.

04/agent next

Recommend the single next actionable move across the workspace.

INITIALIZE THE FRAMEWORK

Give your next session a colleague, not a blank slate.