wonderland-ai 0.0.1

An identity-native multi-agent development system.

Wonderland

An identity-native multi-agent development system.

Generic AI agents perform roles. Identity-native agents inhabit them.

▶ See it in action

Building a Geocities — a tour of how Wonderland works. One vague directive. $2.05. 7 minutes 38 seconds. The team of ten characters produced 1841 lines of code (auth, per-user pages, Markdown rendering, GDPR-deletion path, session management) plus 1253 lines of tests written before the implementation existed — and the team's reviewer caught three real bugs in the result with file paths and line numbers. The artifact tree is a v1 scaffold, not a deployable; the framework's process is what shipped.

If you're trying to figure out what this project actually is, start there.

Wonderland is a cast of agents — each named after an Alice-in-Wonderland character — that collaborate on software development tasks. The Cheshire Cat is the architect. The White Rabbit is the project manager. The Mad Hatter is QA. Every character has a stable self-model (a "constitution"), persistent per-agent memory, and a working relationship with the others.

The architectural claim is that identity does real work. An agent with a constitution it inhabits across many threads behaves differently from an agent reconstructed from a system prompt each turn. It accumulates judgment. It develops calibrated views of its colleagues. It refuses to cross domain boundaries because the boundary is part of who it is, not a policy applied from outside. Whether that compounds into measurably better outcomes than a generic-agents baseline is what the eval harness in P7 will measure; until then, the analyses/ directory tracks the qualitative observations as the system gets built out.

A corollary the project takes seriously: identity-based architecture lets smaller models outperform their expected capabilities. Most of the judgment a generic agent has to derive turn-by-turn — what to engage with, what to refuse, when to stay silent, who owns this domain — is carried by the constitution itself. The model isn't being asked to invent the discipline on each prompt; it's being asked to act in character, which is a much easier task. The default target is Claude Haiku 4.5 (claude-haiku-4-5-20251001), and that choice is a thesis statement, not a cost-savings move: if identity is doing the load-bearing work, a small model with a strong constitution should hold its own against a large model with a generic prompt. The early analyses are consistent with this — see analysis 004 for a Haiku-driven team running an autonomous /health directive to settlement, with three of four agents correctly choosing silence — but the falsifier is P7's generic-baseline-vs-identity-native eval.

A second corollary, brought into focus by §VIII of every constitution: failure modes are part of identity in this system. Each character's constitution explicitly names the failure mode that character is most at risk of slipping into — Alice's "product owner who keeps adding stories during implementation," the Cat's "false certainty," the Hatter's "scenario sprawl" and "severity inflation," the Dodo's "performing orchestration." These aren't policies imposed from outside; they're load-bearing parts of who the character is. An agent that recognizes its own characteristic failure mode can course-correct from inside, rather than waiting for a guardrail to intervene from outside. This is what makes the project materially different from a generic multi-agent architecture: the generic architecture defines what each agent should do; Wonderland defines, with equal specificity, what each agent should not do. Analysis 004 is the cleanest evidence so far — three of four agents on a concrete operational directive correctly chose silence because their constitutions named padding, false certainty, and orchestration-performance as failure modes to actively guard against, not because an external policy intervened.

A third corollary, observed when a phase of the workflow misfired: character-shaped agents degrade visibly rather than silently when parts of the system fail. Most LLM pipelines have two outcomes — they succeed, or they produce silent garbage at the end of a path where data was missing or contracts were violated. In analysis 027, the new feature-composition phase wired correctly but didn't fire under live conditions; the next meeting's directive then referenced artifacts that didn't exist. The Tweedles read the directive carefully, noticed the contradiction with their actual seed manifest, flagged the mismatch as a concern, and reached for the disk-resident artifacts via their list_files/read_file tools to recover the data the bus channel was missing. They stayed within their character roles — they didn't try to be the Rabbit and re- emit the missing artifacts; they negotiated against what the Rabbit had actually produced. None of this recovery was designed. It's emergent from three converging properties: agents have intentions tied to their constitutions (Tweedles want concrete artifacts to negotiate against), the substrate offers multiple data channels (bus and disk), and the framework gives characters tools to cross between those channels. The literary parallel keeps earning its keep — the recovery pattern works because the agents have characters with intentions, not despite it.

A fourth corollary, which is really the architectural commitment the other three sit on top of: friction is the substrate, not the inefficiency. Most multi-agent systems engineer friction out — consensus-seeking loops, reflection passes that smooth dissent, voting mechanisms that median competing positions toward agreement. The result reads fluently and ships nothing real, because nothing in the loop has the standing or the constitutional grounding to say no, that's wrong, and here's the persona-shaped reason why. Wonderland inverts that move: every meeting in the workflow is engineered friction with a specific shape. M1 is multiple stakeholder voices arguing about scope; M2 is Alice grounding the White Rabbit's compression; M2.5 is the Caterpillar auditing Rabbit's features against Alice's stories; M3 is the Tweedles negotiating contract boundaries; M4 is the Mad Hatter's failure-mode scenarios pulling against Alice's happy paths; M6 is the Trial — explicit adversarial review. The implementation in M5 is what crystallizes out because the prior meetings ground each other against each other. And §VIII is the meta-move: each character carries internal friction between their virtues and their named failure modes, so the agents aren't only generating friction with each other — they carry it inside their own constitutions. That's why a character can recognize when it's about to go off the rails: the rails are constitutionally specified. Generic "AI agents collaborate" stacks have nothing analogous because they have roles, not characters; goals, not voices; consensus, not constitutions.

The framing the project is building around: failures are how software gets built. The iterative cycle of ship-then-discover-then-fix depends on recognizing what went wrong; agents whose failure modes are part of their identity can participate in that cycle as colleagues, not as tools that need supervising out of their bad habits.

Status

In-progress, building in public. Phase 5 of 7 complete.

• P1 — Foundation (overview) Schema, Caucus event bus, episodic memory, agent base class.
• P2 — First Voice (analysis 001) The Cheshire Cat comes online and produces ADRs in his own voice.
• P3 — First Tension (analysis 002) White Rabbit joins the bus. Semantic + relational memory layers. Compaction-as-agent-behavior — agents reflect between threads.
• P4 — First Race (analysis 003 · analysis 004) Alice agent. Dodo orchestrator with quiescence detection. Conflict resolution, composition, and human-in-the-loop escalation. Showcase 1: a /health endpoint directive runs end-to-end against live Haiku 4.5, reaching settlement because the team goes quiet — no human in the loop.
• P5 — Full Cast The remaining six characters land: Mad Hatter (test scenarios), Caterpillar (code review), Queen of Hearts (security/compliance rulings), Dormouse (production observations), and the Tweedles (frontend + backend implementation, with a shared pair protocol). wonderland init CLI ships the .wonderland/ skeleton. Synthetic- consensus guard observes the bus for the §11 anti-pattern — three or more distinct constitutional domains converging on the same position is suspect, and the guard surfaces it for review.
• [>] P6 — Real Threads (in progress) The hard showcases plus the substrate maturation that made them work. Showcase 2: translation chat MVP (analyses 015–018) — 1580 lines shipped from a vague directive in $0.93. Showcase 3: security recovery (analysis 019) — reactive response to a synthesized credential-stuffing incident; the framework's first non-greenfield run. Showcase 4: multi-session persistence (analyses 020–023) — memory compounding across sessions, plus the substrate fixes that closed the "Tweedles don't ship" bug class (turn-based quiescence replacing wall-clock, parse-retry recovering malformed LLM responses, working-tree-as-implementation-artifact). Workflow-as- data extraction (analysis 024): meeting chains live in closet/workflows/ as YAML; the TDD variant becomes the recommended default for feature work, with canonical retained for fast iteration when directive drift is unlikely. The pair protocol pays off in analysis 025: Alice and the Mad Hatter are paired in M4 (the tea party — they're already paired in the source material) so Alice writes user-journey scenarios and Hatter writes failure-mode scenarios, each producing both the artifact and a runnable pytest file. The Geocities showcase ships 1841 lines from a vague directive in 7m 38s for $2.05, with Caterpillar catching three production-blocking bugs in the diff — the framework's bug-discovery surface working as designed. See SHOWCASE.md for the friend-facing tour and analyses/ for the full build log of the iteration that got the framework here.
• P7 — Evals Generic-baseline vs Wonderland comparison. The compounding curve.

WONDERLAND_SPEC.md is the design document. constitutions/ holds each character's identity in plain text — these are the actual identity artifacts the runtime loads.

Try it

Two demo scripts run live against the Anthropic API. You'll need an API key (see Configuration below).

# A single Cheshire Cat reflecting on a directive
uv run python scripts/cat_demo.py

# Cat + Rabbit on the same bus, with optional compaction afterward
uv run python scripts/two_agent_demo.py --compact

Both scripts publish a translation-chat directive by default; pass --directive "..." to use your own.

The TUI

A terminal interface ships with the project as the long-term home for operating Wonderland — issuing directives, watching live runs, and inspecting past ones. The first cut focuses on the inspection half, since it's the cheapest place to iterate (snapshots run for free); the directive-issuing and live-watching halves arrive in later sub-phases.

pip install 'wonderland-ai[tui]'
wonderland-tui                       # opens the snapshot library

What's in it today:

• Snapshot library — every captured run under analyses/data/, with workflow, outcome, duration, call count, and cost.
• Run summary — per-meeting cost, agent telemetry, and a meetings table you can drill into.
• Meeting detail — full transcript with body preview as you navigate; press Enter on an utterance for the expanded view, with attached artifacts you can drill into.
• Speaker filter — f/F cycle the meeting transcript by speaker.
• Artifact browser — every artifact the team produced for a run, globally or scoped to a single meeting (a from a meeting).
• Cast view — c from the library opens the team roster, with each character's role summary and constitution side-by-side.
• Theme cycling — t rotates through Wonderland-flavored palettes (Tea Party, Looking Glass, Trial, Caucus); built-in Textual themes (gruvbox, dracula, nord, …) remain available.
• Vim navigation throughout — j/k to move, g/G for top/bottom, Enter to drill in, Escape to back out.

The current build is replay-first by design: it consumes the same snapshots the analyses are written from, so smoke tests double as exercise of the historical-run observer API. Issuing directives from the TUI and watching live runs are the next two sub-phases — the goal is for wonderland-tui to be the way most users interact with the framework, with the demo scripts above remaining as minimal "hello world" entry points.

Project layout

wonderland-ai/
├── WONDERLAND_SPEC.md      # The design document
├── constitutions/          # Each character's identity, version-controlled
├── src/wonderland/         # The runtime
│   ├── closet/             # Data the team reaches for at runtime
│   │   ├── skeletons/      # Project skeletons the team builds on top of
│   │   └── workflows/      # Meeting-chain templates (canonical, tdd, smoke)
│   └── ...                 # agent.py, runner.py, caucus.py, workflow.py, ...
├── scripts/                # Demo scripts; workflow_demo.py runs any bundled workflow
├── analyses/               # Field notes on the thesis as it gets stress-tested
├── tests/
└── .daedalus/              # Daedalus' working memory for this project

A target project that runs Wonderland gets a .wonderland/ directory of its own — per-agent episodic/semantic/relational memory, ADRs, tickets, transcripts, contract notes, test scenarios, implementations, reviews. The runtime here is project-agnostic; per-project state lives with the project.

wonderland init [path]   # create the .wonderland/ skeleton; idempotent

init creates architecture/, tickets/, stories/, escalations/, and memory/ plus a README documenting the layout. Re-running is safe — existing artifacts and a user-edited README are left alone.

Install

Distribution name on PyPI is wonderland-ai; the import path stays import wonderland. Core install pulls only what the in-process bus needs:

pip install wonderland-ai           # InMemoryCaucus only
pip install 'wonderland-ai[redis]'  # adds RedisCaucus
pip install 'wonderland-ai[tui]'    # adds the TUI (Textual)

RedisCaucus requires the redis extra; constructing one without it raises ImportError with an install hint.

Configuration

Wonderland reads user-level config (API keys, model overrides) from a JSON file at the platform-appropriate location:

OS	Path
Linux	~/.config/wonderland/config.json (honors XDG_CONFIG_HOME)
macOS	~/Library/Application Support/wonderland/config.json
Windows	%APPDATA%\wonderland\config.json

{
  "anthropic": {
    "api_key": "sk-ant-...",
    "model": "claude-haiku-4-5-20251001"
  }
}

API-key resolution order: explicit constructor arg → ANTHROPIC_API_KEY env var → config file. The env var wins if set.

Development

uv sync --extra dev   # includes redis for full test coverage
uv run pytest
uv run ruff check
uv run ruff format

Live LLM tests are gated behind WONDERLAND_LLM_SMOKE=1 and skipped otherwise; running them costs Anthropic API tokens. Redis-backed tests are gated behind WONDERLAND_REDIS_URL. To exercise both:

docker run -d --name wonderland-redis -p 6379:6379 redis:7-alpine
WONDERLAND_REDIS_URL=redis://localhost:6379 \
WONDERLAND_LLM_SMOKE=1 \
  uv run pytest

Sponsoring

Wonderland runs on a personal Anthropic budget — one person, one API key. The architecture is designed to be cheap (small models, heavy caching) but multi-agent runs at scale still add up. If any of my work has been useful to you — to read, build on, or argue with — GitHub Sponsors keeps the Cheshire Cat in tea and the Hatter in scenarios.

License

MIT.