PRODUCT — CODIFY YOUR BUSINESS

A team of small specialist agents.
One per function you already do.

PaperAgents is the entrepreneur-facing surface of the QuKaiZen stack — declare a team in TOML, hand it the work you keep repeating, and let a watcher reconcile drift while you sleep. Underneath, it's a configuration-first orchestrator built on top of the open-source Paperclip AI platform. ARAIL is the lab. PaperAgents is the company.

STACK — READ BOTTOM UP01 — YOUR TEAMyou compose · specialists wrapped around real functionsSales Agentlead triageSupport Agentticket loopOps AgentschedulingBooks Agentreconcile+ yours02 — PAPERAGENTSplatform · TOML teams · watcher · per-team RAG · reconciliationWatcher AgentcontinuousDeclarative TeamsTOML / CRDPer-Team RAGembeddedDrift Reconcilerto specIdempotent Applyrollback safeEvent Relaycached · observable03 — PAPERCLIP AIMIT foundation · plugin host · event bus · governancePlugin Hostruntime-agnosticEvent BusheartbeatsGovernancebudgetsMulti-TenantisolatedMODEL TIERSescalate when neededARAIL Labtier 3 · researchoriginal researchescalateFrontiertier 2 · reasoningdeep reasoningescalateTeamtier 1 · multi-modalsmall · purpose-builteasy to load · defaultspend rises with leveragemost work stays at tier 1You composePaperAgents ownsPaperclip AI providesModel tiers (escalation)

THE WEDGE — DECLARATIVE SOURCE OF TRUTH

Everything flows from TOML. Diagrams, audit trails, cost reports, escalation policies — all auto-generated from your declared truth.

The running daemon is a deterministic projection of applied manifests. No team structure, no routing rule, no budget envelope exists anywhere except as a consequence of the TOML you committed. That's the difference between paperagents and the orchestrators and agent frameworks it competes with.

TEMPORAL / AIRFLOW

Imperative code

Workflows are compiled binaries. Diagrams are drawn by hand and drift the moment someone touches the code. “What does this DAG actually do?” means reading the source.

LANGCHAIN / LLAMAINDEX

Embedded config

Agent logic is Python. Escalation is hardcoded if/else. Governance, budget, RBAC, audit — application-level, custom per project. There is no source of truth — there's a repo.

PAPERAGENTS

Declarative TOML

One file declares teams, agents, tiers, budgets, escalation predicates, workflows. Runtime, diagrams, audit, costs — all projections. Git-versioned. Idempotent apply. No drift, by construction.

TOML → EVERYTHING

One source of truth fans out into runtime state, generated diagrams, attributed cost reports, and a compliance audit trail. Edit the file, re-apply, every downstream artifact moves with it.

Source of truth flow

DEEPER — HOW IT WORKS

WATCHER AGENT

Continuous monitoring loop

A long-running agent that watches the Paperclip ecosystem, upstream changes, and the team's own state. Surfaces drift the moment desired-state and observed-state disagree. The watcher never sleeps; you do.

DECLARATIVE TEAMS

TOML-over-YAML · CRD model

A team is a config file, not a script. Human-readable TOML declares which agents the team has, what work they own, and the budget envelope. Idempotent apply, graceful drain, and rollback are first-class. You git-commit your org chart.

EMBEDDED PER-TEAM RAG

Each team gets its own memory

Multi-tenant by default. Sales has its own corpus. Support has its own. Books has its own. No cross-bleed, no shared context window, no accidental customer data in the wrong loop. Isolation is a contract, not a guideline.

BUILT ON PAPERCLIP AI

MIT · open-source foundation

Paperclip AI is the orchestration substrate — plugin host, event bus, governance, multi-tenant isolation. Any agent, any runtime, one org chart. PaperAgents adds the configuration-first platform layer so a non-engineer can run a team. paperclipai/paperclip ↗

SIBLING TO ARAIL

Same stack · two audiences

ARAIL is the lab — for researchers and builders running experiments. PaperAgents is the company — for entrepreneurs codifying repeatable work into a team. Neither is forked from the other; they target different rooms in the same building.

DRIFT RECONCILIATION

Desired state is the contract

The platform compares what the TOML says should be running against what is actually running, and closes the gap. Restart, drain, re-apply, roll back — all observable, all idempotent. The org chart converges back to the file you committed.

ROADMAP — HARNESS ENGINEERING

Four features. Four weeks of critical path. Each one is clean, secure, standard. None duplicate Paperclip AI.

Hot reload for DX. Cost chargeback for the operator who actually pays the bill. Durable workflows for work that takes hours or days. Auto-generated diagrams so the org chart never lies. Click any card to see the deliverables, the rationale, and the CLI surface.

PROOF — VS THE FIELD

What paperagents is not.

Temporal is a workflow engine. paperagents is the governance scaffolding that includes one. Different scope, different audience.

AspectTemporal.iopaperagents
Workflow definitionGo / Java / Python codeDeclarative TOML
DurabilityEvent-sourced tlogCheckpoint-first
EscalationNot a conceptCost-aware tier routing
Cost trackingNot built-inFirst-class, with alerts
Schema validationNoEvery step boundary
Config managementCodeGit-versioned TOML

DIAGRAMS — RENDERED LIVE

Six views of the same TOML.

These diagrams are reference renderings of the shape pactl graph emits at runtime — same Mermaid source, same projection rules. Copy any of them; paste into mermaid.live to inspect.

Governance layer (paperagents) sits on the open-source Paperclip AI runtime. Clients apply TOML; escalation routes tasks by cost-aware tier.

paperagents on top of Paperclip AI

WHY IT MATTERS

DECLARATIVE = OBSERVABLE

The truth is a query, not a code-read.

“What is the escalation policy?” → pactl get agents. “Show me the org chart” → pactl graph --org. No guessing. No archaeology. Just facts, in TOML.

SOURCE OF TRUTH = NO DRIFT

The diagram is never stale.

Diagrams are auto-generated. Audit trails follow from declared config. Cost reports reflect declared escalation. One edit, everything moves with it. Drift is a category error, not a bug.

PLATFORM-ENFORCED = SECURE

Guard rails before execution, not after.

Budget limits enforced before the call goes out. Secrets never touch logs. Circuit breakers fail fast. Schema validation at every step boundary. Compliance falls out of declared truth.

MODEL TIERS — ESCALATE WHEN NEEDED

TIER 1

Team agents do the routine work with multi-modal tools (text, vision, voice…).

TIER 2

Frontier models can be leveraged when an agent hits a problem that needs deep reasoning — and is the default for the CEO.

TIER 3

ARAIL research lab when the question needs actual research.

Most work happens at tier 1 on small, purpose-built, easy-to-load models. Compute follows leverage — frontier reasoning and original research are reserved for the questions that actually need them.

TIER 1 — TEAM AGENTS · DEFAULT

Small, purpose-built, multi-modal

Sales, support, ops, books — each picks the modality that fits the task. Text and code for paperwork. Vision for screenshots and invoices. Voice and audio for calls. One team, the right tool per role — running on small models that load fast and stay cheap.

TIER 2 — FRONTIER · CEO DEFAULT

Deep reasoning on demand

The default model class for the CEO agent — and the escalation target whenever a team agent hits a problem its small model can't crack. Claude, GPT-5, Gemini, Llama 405B. Used by default at the top, surgically below it.

TIER 3 — ARAIL LAB · RESEARCH

Original research, not just thinking

Some questions need an answer that doesn't exist yet. The CEO agent — or any escalated request — can dispatch to ARAIL, where autoresearch agents actually investigate, curate sources, and report back. The company's research arm.

Forward-looking: as the QuKaiZen Nucleus pipeline mints sealed Super Skill Models, those slot directly into tier 1 — domain-specialised, cryptographically signed, runnable on commodity hardware. Same pattern, stronger small models.

Learn →