Maxey0

Route a task

The decision this system exists to make: does a hardened loop already cover this work — with its partition already configured — or must one be assembled from skills, then agents?

What is loaded

Sources

How this works

1

Concepts route

A task is scored against Maxey0's 16 concept tag vocabularies. The winning concepts decide which loops are even candidates.

2

Loops carry their partition

Each loop in the library was hardened by being built against a real ContextWindow: regions created, roles bound, every negative control actually attempted and actually refused.

3

SCWs enforce it

A bound role reads its own region and whatever it holds an explicit, expiring grant for. Anything else raises a refusal — that is the wall, and it is in the runtime, not in a prompt.

Concept → Skill → Agent → Dispatch

Creating an SCW here goes through the same hierarchy as everything else in Maxey0 — pick what's doing the work before there's anywhere for it to write. Every region created is real; every dispatch is a real write() against a real, enforced token budget.

concept

skill

agent

target scw

label

count

region type

token ceiling

prompt

—

drag a file here to write it into the target scw

anthropic api key

checking…

Observation grid

Concepts — click one to see the loops that touch it

Agentic loop library

Status is evidence, not decoration. validated built clean against a live window; partial has a real, named problem; draft-unexecuted has never been run — its own README says so.

Agents

New loop

Drag agents in from the left, or click them. Drag a stage to reorder. Each stage becomes one role bound to its own private scratchpad, publishing to one declared handoff region — so stage N reads stage N−1's output, never its reasoning.

Drop agents here to build the pipeline

Regions

view as

Pick a role to read the window through its scope. Regions outside that role's closure come back refused, with the runtime's own reason — not hidden by the UI.

Trace

live

Semantic plane — three hierarchy levels, one anchor field

concepts skills agents edges

Containment

Two of the four assertions attempt a real read through the real authorization path, so a refusal is an event the ledger records and replay reproduces. The other two compute over the region graph. Every row says which you are looking at, because a set computation is a weaker claim than an attempt the runtime actually refused.

Nothing loaded.

Isolation level

A level is earned from evidence, never asserted. Declare L3 and get L1 and nothing else in the stack would notice.

Nothing loaded.

Event stream

contained is three-valued. Null means nothing was attempted, which establishes nothing either way — it is not a clean result.

Nothing loaded.

Gate

Every tool call a delegated agent made, attributed to the SCW role it was bound to. This is the stream the runtime's own log cannot contain: it records what the host did to the window, while an agent dispatched as a subagent works in a context the window has no address for. The gate stands at the tool call, which is the one place that context has to ask for something.

auto (2s) —

Isolation level

A level is earned from evidence, never asserted. Declare L3 and run L1 — three roles dispatched into one shared context, the partition existing only in the prompt — and nothing else in the stack would notice.

declared

Containment

No gate activity recorded yet.

Per role — context in, execution out

What each role was handed, what it then did, and a finding where those two disagree. A role given context that made no observed call, and one that acted with no recorded transfer, are both worth knowing.

Route a task

What is loaded

Sources

How this works

Concepts route

Loops carry their partition

SCWs enforce it

Concept → Skill → Agent → Dispatch

Observation grid

Concepts — click one to see the loops that touch it

Agentic loop library

Agents

New loop

Regions

Contents

Trace

Semantic plane — three hierarchy levels, one anchor field

Containment

Isolation level

Event stream

Gate

Isolation level

Containment

Per role — context in, execution out

Activity