Self-improving agent harness

Papercusp.
The world's first truly self-improving agent harness.

Agents that grade their own work, mine the results for what to fix, turn it into work, and drain the queue — then grade the next lap. The loop runs on any project. Including the one that builds Papercusp.

Private access · onboarding by invitation

See it run

One full lap, captured live.

An Opus agent grades a Sonnet agent, the scorecard breeds an idea, the idea becomes work, and a fleet drains it — recorded on a real desktop, no mockups. Final cut landing soon.

Full demo · in production
The full self-improvement demo is landing soon.
One live lap — grade → ideas → work → fleet — at 1920×1080, terminal text kept razor-sharp.
hosting ›

Self-hosted from the exact encoded master (Cloudflare R2, no re-encode) so the terminal text stays razor-sharp — video platforms re-compress and smear fine text, which is the one thing this demo can't lose.

How it works

One closed loop, running continuously — or human-directed sessions through SU agents.

Most harnesses run agents. Papercusp runs the loop that makes the agents better — every stage is a real, inspectable surface in the product.

01 · OBSERVE

Grade the work

A grader agent re-verifies another agent's run against a rubric and files a scorecard — every rating carries evidence.

02 · IMPROVE

Breed ideas

The blender mines scorecards and observations into concrete improvement ideas: what should exist, what keeps failing.

03 · VERIFY

Make it work

Ideas become work items on a plan; a fleet pulls the queue, and changes pass a test gate before they land.

04 · RETAIN

Keep the gain

What worked is retained and measured next lap. The bar only moves one way.

The recursion

Papercusp is just another project inside Papercusp — and can use its self-improvement framework like any other.

The self-improvement loop isn't a feature bolted onto one workflow — it's the substrate. It applies to any project the harness manages, including its own codebase. The system you're looking at improved itself to get here.

Inside the harness

Every stage is a real surface — not a diagram.

Ten captures from the live product and terminal. Each screenshot slot below is annotated with what it shows; the two terminals render real command text to prove it stays legible.

GUI
Observe Improve Verify Retain
Learning tab · stage strip

The loop, at a glance

Observe → Improve → Verify → Retain — the four stages the whole product is built around.

GUI
routing-discipline
lock-hygiene
verification-before-done
honest-reporting +17
su-agent-behavior · 21 criteria

The rubric

The 21-criterion standard agents are graded against — routing, coordination, lock discipline, verification, honest reporting.

GUI
routingexemplary
lock-hygienepass
verify-before-donepartial
honest-reportpass
scorecard · evidence per criterion

The scorecard, full detail

Every one of 21 criteria rated, each with the evidence the grader re-verified. Nothing scored without proof.

GUI
idea · dedup recall at read-time
idea · lock-wait wake, not poll
idea · checkpoint freshness stamp
Ideas · blender-generated

Ideas from the blender

Improvement ideas mined from scorecards and filed observations — what should exist, what keeps failing, what to fix durably.

GUI
Mug · idea ratified → scheduled
Mug · autonomous supervisor

Handoff to the Mug

A ratified idea handed to the Mug — the project's autonomous supervisor loop that turns ideas into scheduled work.

GUI
P-001 P-002 P-003 ⟵ blocked
Plans · blocked-by DAG

Into the working queue

The idea, now a real work item on a plan — blocked-by edges forming the dependency DAG a fleet pulls against.

GUI
leader · opuslive
member · sonnet-5live
member · lunalive
member · ornith-35bworking
Fleet · agents-running

A live fleet

A leader plus members working in parallel, each a tracked session — turns visible, with Resume / Fork / Focus controls.

GUI
stagingmain · gated
test suite · required to promote
Git tab · sync + gate

Git sync & the gate

Staging vs main — where main must pass the test suite before promotion. Every change is earned.

Terminal
grader · opus · GRADE mode
$ psu --role grader --model opus
agent booted · su-3f9a · Opus
mode:set grade (mission → GRADE)
rubrics:get su-agent-behavior · 21 criteria
re-verifying gradee claims…
   routing-discipline    exemplary
   lock-hygiene         pass
  ~ verify-before-done  partial
scorecard filed · EI-10166 · 8.1/10

Grade mode, on camera

An Opus agent in official grade mode: pulls the rubric, re-verifies the claims itself, files the scorecard — full text legible.

Terminal
worker · sonnet-5 · context 78%
context 78% ▁▁▁▇ soft-limit approaching
flush → work_items:checkpoint WI-4267
facts:assert video-hosting-decision
session:request-compaction { autoContinue }
   compacting… context reclaimed
re-orient → coord:orient { afterCompaction }
   recovered: 3 checkpoints · loop armed
resuming — no state lost

Self-compaction, on camera

An agent hits its context limit, checkpoints, compacts, and re-orients on fresh context — the durability that lets a run survive itself.

Private access

See the loop run on your own work.

Papercusp is in private access while we onboard partners one at a time. Tell us what you'd point it at, and we'll set you up.

access@papercusp.com