drag the concepts · watch the text respond · zero DOM reads
Lines0
Reflow0ms
DOM reads0
FPS60
Columns0

what you get

a complete agent context layer. works with any harness. portable across tools, persistent across sessions.

memory

every decision, discovery, and learned pattern persists. graph edges, embeddings, hybrid search. your agents remember what worked.

searchable knowledge graph

context hub

MCP server your agents connect to. semantic search across your entire project. works with any harness.

MCP compatible

training loop

every action produces a (state, action, outcome) tuple. the policy head trains on your data. learns YOUR patterns.

RL from real outcomes

autonomous agents

agents run overnight. try changes, eval results, keep what improves, revert what doesn't. wake up to PRs.

Karpathy autoresearch

build evals

write a spec, the eval checks if it's built. agents iterate from 0 to 100 percent autonomously.

spec is the eval

journals

every session writes structured entries. decisions, features, discoveries. future sessions start with full context.

persistent context

world model

tracks state transitions, predicts outcomes, detects when assumptions break. agents reason, not just act.

predictive scheduling

agent mesh

P2P network for agent coordination. zero-config discovery, encrypted messaging, pub/sub.

Subway P2P

how it works

10ET sits between your agents and your project. it captures context, accumulates training data, and improves the policy — automatically.

01

your agent works

using any harness — Pi, Claude Code, Cursor, a custom script. 10ET provides context via MCP. the agent reads memory, past decisions, and experiment history before making changes.

02

every action produces training data

what was the state? what action was taken? what was the outcome? this tuple gets written to the training buffer. decisions get journaled. new knowledge gets indexed into memory.

03

the policy improves overnight

nightly: the policy head retrains on accumulated tuples. eval scripts measure real metrics. agents try changes, keep what improves, revert what doesn't. PRs get created automatically.

04

next session starts with everything

memory of what worked. trained preferences. past experiment history. the agent doesn't start from zero — it starts from the accumulated intelligence of every session before it.

# the loop in practice $ jfl peter agent memory-recall -r 5 Baseline: 0.429 Round 1: 0.571 (+0.14) KEPT Round 2: 0.643 (+0.07) KEPT Round 3: 0.714 (+0.07) KEPT PR created: github.com/.../pull/94 # next night: starts from 0.714, not 0.429 # the world compounds

self-organizing systems don't need a conductor. they need shared context and a reason to coordinate.

this is how worlds compound

every session makes the next one better.

stop starting from zero. give your agents memory, training, and a world model that compounds.