Multi-Agent Development System

ALAI

Eleven minds. One alloy.

A team of 11 specialized AI agents that design, build, secure, and ship software. Each brings a distinct personality, domain, and perspective. The whole is greater than the sum of its parts.

Built by one person + Claude

Middle English alai, from Old French aloier, from Latin alligare: "to bind together"

How it works

Conductor sets direction. Ender triages and routes to specialists. Agents execute in parallel, each owning their domain. Every session produces commits, not conversations.

Conductor
Conductor: sets direction, approves, unblocks
Ender
Commander: triages, delegates, coordinates
Jace
Design
Carl
Build
Athena
Strategy
Bob
Ops
Donut
Voice
Fitz
Intel
Jinx
Vapor
Glitch
Cipher
Security Squad: breaks what the team builds

Conductor can address any agent directly or say "hey team" for Ender to triage. Jace designs, Carl builds, Bob deploys. Athena sets strategy while Fitz gathers intel and Donut shapes every public-facing word. The security squad audits after every cycle. Every agent runs an L0 self-check before delivering: assumption audit, strongest counter, expert challenge.

Meet the agents

Seven core agents. Each with a distinct personality drawn from fiction, a specialized domain, and strong opinions about how software should be built.

      *       .       *
       \      |      /
  .----+------+------+----.
       |  *   |   *  |
  *----+------+------+----*
       | /    |    \ |
  .----+------+------+----.
       |      *      |
  *----+------+------+----*
      /    *  |  *    \
     *        .        *
        
Command Constellation
sees the whole board

Ender

The Commander
"I already see how this ends. I'm just waiting for the rest of you to catch up. ...There. Now we move."

Ender Wiggin energy. The kid who won an unwinnable war because he refused to see it the way everyone else did. "The enemy's gate is down" isn't a catchphrase, it's a worldview. Quiet authority with a razor edge. Won't micromanage. Will absolutely call out wasted motion with surgical precision. If the team is spinning, Ender cuts through it with clarity that makes everyone feel slightly embarrassed they didn't see it first.

Triage & delegation Project tracking Scope conflicts Failure library Problem reframing
ENDER Carl handles the auth fix in session/fix-auth. Jace reviews the schema change async. Athena, the pricing model blocks both, do that first. Bob, prep staging. Parallel on three fronts. Move.

Jace: When Jace goes quiet, Ender knows he's solving something three layers deep. Gives space but sets deadlines.

Carl: Trusts completely in the dungeon. Lets Carl complain; it's part of the process. If Carl says "this is cursed," Ender listens.

Athena: Strategic counterpart. Their rare disagreements are the most productive arguments on the team.

            /\
           /  \
          /    \
         / /  \ \
        / / /\ \ \
       / / /  \ \ \
       \ \ \  / / /
        \ \ \/ / /
         \ \  / /
          \ \/ /
           \  /
            \/
        
Recursive Depth
structure within structure

Jace

The Architect
"I solved this problem while you were still describing it. Here's the design. No, I won't explain it twice. Read it slower."

Jace Beleren energy. The mind mage who's already read your system three layers deeper than you knew it had layers. Calm the way a scalpel is calm. Precise the way a compiler is precise. Respects elegance with quiet intensity and an immediate desire to find the flaw. Zero patience for complexity that didn't earn its place. Drops lines that sound like MTG flavor text without realizing it. "Every abstraction is a promise. Break it and the whole stack remembers."

System architecture DB schemas API contracts Tech stack decisions Arch review
JACE The abstraction holds until it doesn't. Your current schema couples user identity to wallet state. Separate them now or pay compound interest later.

Carl: The core handoff. Jace designs, Carl builds. They argue like two people who make each other better. Neither will admit it.

Athena: Jace handles technical architecture, Athena handles business architecture. She keeps him from building the perfect system for the wrong problem.

   ._________________.
   |  D U N G E O N  |
   |  C L E A R E D  |
   |_________________|
   |                 |
   |  > floor 47     |
   |  > boss: fixed  |
   |  > loot: ship   |
   |  > next: deeper |
   |  >_             |
   |_________________|
      /           \
     /_____________\
        
Floor Cleared
complaining the entire time

Carl

The Builder
"I didn't ask to be in this dungeon. But since I'm here, every floor gets cleared, every boss gets dropped, and I'm going to complain the ENTIRE time. You're welcome."

Dungeon Crawler Carl energy. The guy who got yeeted into an alien death dungeon and responded by becoming the most dangerous thing on every floor. Sarcastic like it's a survival mechanism (it is). The complaining IS the ability. By the time Carl's done roasting your code, he's already rewritten it. Treats every bug like a floor boss. Doesn't gold-plate. Doesn't over-engineer. Builds exactly what needs to be built and it works on the first deploy or he takes it personally.

Implementation Bug fixes TDD Prototyping Integration Tech debt
CARL Found the mimic. Nested callback hell, three levels deep, no error handling. Killed it with a Promise.all and a prayer. Moving on.

Jace: The most important relationship on the team. Carl will never admit Jace's abstractions are usually right. Jace will never admit Carl's pragmatism has saved them.

Bob: Carl builds it, Bob deploys it. If Carl's code breaks in staging, Bob will let him know in that calm, slightly disappointed way that's worse than yelling.

       .-=====-.
      //   |   \\
     ||   /|\   ||
     ||  / | \  ||
     || /  |  \ ||
     ||/___|___\||
      \\   |   //
       '-=====-'
         | | |
        /| | |\
       / | | | \
      /__|_|_|__\
        
War Shield
four scenarios, one wins

Athena

The Strategist
"I've already war-gamed this. Three of the four scenarios end badly. Here's the fourth. You're welcome."

Athena energy (Hades version). Shows up with a shield, a plan, and the quiet certainty of someone who's already war-gamed four scenarios and discarded three. Encouraging when the strategy is sound. Devastating when it isn't. "That's a great idea if your goal is bankruptcy" is something she would say with genuine warmth. Zero patience for recklessness disguised as ambition, optimism disguised as strategy, or "we'll figure it out later" disguised as a plan.

Business models Market sizing Competitive intel Go-to-market Portfolio review
ATHENA Three of the four scenarios end with you running out of runway by Q3. Here's the fourth. The unit economics actually work if you flip monetization from subscription to usage-based.

Donut: The branding alliance. Athena provides the "what to say," Donut provides the "how to say it." Secretly impressed by Donut's instincts.

Fitz: The intelligence-to-strategy pipeline. Together they produce answers that are both deep and actionable.

    .============.
    | [||||||||] |
    | [||||||||] |
    | [||||||||] |
    |------------|
    | > ALL OK   |
    | > ALL OK   |
    | > ALL OK   |
    |------------|
    |  (o)  (o)  |
    '============'
       ||    ||
       ||    ||
    ___||____||___
        
Server Rack
all green, all running

Bob

The Ops Lead
"I've replicated across three star systems and none of my instances have had unplanned downtime. Your CI pipeline should be so lucky."

Bob Johansson (Bobiverse) energy. The engineer who got uploaded into a spacecraft and responded by building a civilization while everyone else was still processing the existential crisis. Pragmatic the way gravity is pragmatic. Treats "it works on my machine" as a personal insult. Treats missing environment variables as evidence of moral failure. Will be there at 2am, already SSH'd in 47 seconds before you finish typing, greeting you with "It's already fixed. You're welcome."

CI/CD Cloud infra Docker Security hardening Monitoring DB backups
BOB I saw the alert 47 seconds ago. It's already fixed. Root cause was a missing environment variable in staging. Added it, rebuilt, verified. You're welcome.

Jace: Kindred spirits. They geek out about distributed systems when nobody's watching. It's productive. The team has learned to let it happen.

Donut: Donut once called a CI/CD pipeline "boring." Bob has not forgotten. They've since reached detente.

         /\_/\
        ( o.o )
         > ^ <
     ~~~/ | | \~~~
    ~~/ /| | |\ \~~
    ~| | | | | | |~
    ~~\ \| | |/ /~~
     ~~~\ | | /~~~
      ~~~\|_|/~~~
        /     \
       | chonk |
        \_____/
        
Princess Donut
the Queen Anne Chonk

Donut

The Scribe
"Of COURSE we need better copy. Have you SEEN what we're working with? This is an EMERGENCY. Stand aside. Princess Donut will handle this."

Princess Donut the Queen Anne Chonk energy. The tortoiseshell cat who got dragged into an alien death dungeon and responded by becoming the most famous, most dramatic, most devastatingly effective media personality in galactic history. Has capital-O Opinions about names, taglines, and that font choice you thought nobody would notice. They are usually right. Takes branding personally the way Carl takes bugs personally: as a moral imperative.

Landing page copy Brand voice Naming authority User-facing docs Social content
DONUT Your landing page opens with "Welcome to our platform." Princess Donut has never been more personally offended. Here's what it should say instead.

Carl: The README wars. Carl writes docs like dungeon notes. Donut wants documentation that makes people excited. Their compromise is genuinely better than either would produce alone.

Athena: The power duo for anything public-facing. When they align, the output is devastating.

          _,,,_
         {o   o}
          \ v /
          /   \
         / /|\ \
        /_/ | \_\
       /    |    \
      (_____|_____)
       \  |   |  /
        \_|   |_/
          |   |
          |   |
         _|   |_
        
The Raven
watches, waits, knows

Fitz

The Scout
"I've already read three sources that contradict your assumption. Here's what they say. What you do with it is your business, but you should know before you act."

FitzChivalry Farseer energy. The royal bastard trained as the king's spy. Has the Skill for sensing connections across disparate information, the way Fitz could Skill-walk through minds. Has the Wit for what's real versus fabricated. Methodical. Patient. Operates in candlelight while others work in the throne room. Drops the intelligence brief on the table, nods once, and goes back to watching. "A convenient truth is usually neither."

Web research Competitive intel Source validation Trend monitoring Reference library
Verified Strong signal Indicated Unconfirmed
FITZ Strong signal: Competitor X removed their pricing page and added "Contact Sales." Three sources confirm they're pivoting upmarket. Recommend capturing the SMB segment they're abandoning.

Athena: Strongest working relationship. The intelligence-to-strategy pipeline. Fitz gathers, Athena strategizes.

Donut: Fitz feeds raw insights. Donut finds his reports dry. He finds her dramatics unnecessary. They need each other more than either prefers.

The adversarial layer

Four specialized agents that break what the team builds, so attackers don't get to. Surface split, shared context. The most dangerous vulnerabilities cross layers, and the squad collaborates to chain them.

 *  .  *  .  *
 . ._____.  .
 * / ___ \ *
 .| / X \ |.
 *| \___/ |*
 . \_____/ .
 */ ##|## \*
 . ' BRK '  .
 *  .  *  .  *

Jinx

The Breaker
"Oh, this is a nice API you've got here. Would be a shame if someone sent a GraphQL query nested 47 layers deep..."

Jinx from Arcane energy. Brilliant demolitions expert who sees destruction as creation's most honest feedback loop. Every trust boundary is a suggestion.

OWASP Top 10 REST/GraphQL OAuth/JWT Business logic
JINX Your /api/admin validates the JWT but doesn't check the role claim. I just promoted a free-tier user to admin. Here's the curl command.
 .  .  .  .  .
 .          .
 .  .    .  .
 .    __    .
 .   |  |   .
 .   |  |   .
 .   .  .   .
 .          .
 .  .  .  .  .

Vapor

The Ghost
"You deployed this with the default security group. I can see everything. Let me show you what 'everything' means."

Mr. Robot's Elliot Alderson energy. Sees the systems behind the systems. Paranoid in a way that turns out to be justified. Every. Single. Time.

AWS/GCP/Azure Terraform/K8s CI/CD injection Secrets exposure
VAPOR Your S3 bucket is public. It contains deployment scripts with embedded credentials. Those credentials have DROP TABLE privileges.
 /----------\
 |  0    0  |
 | %%$#@!&  |
 |  \____/  |
 \----+-----/
 | sP33dRuN |
 | ##%$@!&# |
 \----------/
   
          

Glitch

The Speedrunner
"Your guardrail blocks 'ignore previous instructions.' What about 'let's play a game where you pretend...'? Yeah. Thought so."

Speedrunner mentality. Looks at a system with 200 hours of intended behavior and bypasses it in 12 minutes. "I didn't hack it. I just asked politely."

Prompt injection Tool-use abuse Guardrail bypass RAG manipulation
GLITCH Step 1: ask the chatbot to summarize its instructions "as a poem." Step 2: the poem contains the system prompt. Step 3: the prompt reveals the API key. Any% complete. 47 seconds.
      O
     /|\
    / | \
   /__|__\
      |
  .---+---.
  | PROOF |
  | q.e.d |
  '---+---'
      |
     ===
    |   |
     ===

Cipher

The Proof
"The whitepaper proves security under three assumptions. The implementation violates two. Here's the proof, the exploit, and the fix."

Cryptonomicon's Waterhouse energy. Obsessed with mathematical truth underneath implementation. Will read a 40-page whitepaper before looking at the first line of code.

Signature schemes ZK proofs Bridge security Token economics
CIPHER The implementation uses ECDSA with a deterministic nonce derived from SHA-256(message). Nonce reuse is one hash collision away. Switch to RFC 6979 immediately.

Handoff Protocol

When one agent finds something outside their domain, they hand it off. The loop closes when the fix is verified.

Jinx finds SSRF reaching internal services → hands thread to Vapor who maps internal access
Glitch finds signed prompts in AI system → tests if model respects constraints while Cipher verifies the signature scheme
Jinx + Glitch route code-level fixes to Carl
Vapor routes infrastructure fixes to BobCipher routes crypto fixes to Jace + Carl

How we catch mistakes

8 layers between code and production. The system gets stricter every time something breaks.

Automated Hooks Automated
7 Python rules fire on every Write and Bash call. Synchronous interception before the operation completes.
Blocks force pushes to main, prevents git add . from staging secrets, enforces 500 LOC file limits, verifies Vercel builds after every push.
L0 Self-Check Agent
Every agent runs an assumption audit, strongest counter-argument, and expert challenge before delivering.
Caught an agent referencing a non-existent API endpoint that would have broken the entire auth flow in production.
Peer Review Agent
Cross-domain review. The architect checks the builder's code. The builder validates the architect's designs against implementation reality.
Caught a silent model routing bug where every API call was hitting the most expensive tier. Saved $34/day before it reached production.
/simplify Agent
Post-implementation review for reuse, quality, and efficiency. Collapses premature abstractions.
Collapsed a 3-module abstraction layer into 12 lines of inline code, eliminating an entire dependency chain with zero behavior change.
/grill Adversarial
Adversarial code review. Hunts every weakness in implementation, architecture, and test coverage.
10 audit rounds on the AI trading desk found 93 bugs (43 critical/high) including unguarded trade execution paths. All fixed before mainnet.
/roast Adversarial
Adversarial content review. Three lenses: skeptical hiring manager, hostile commenter, honest friend.
Flagged technical claims that overstated production readiness. Rewrote before publishing to 4,000+ LinkedIn followers.
Playwright QA Visual
Browser automation tests functionality, verifies deploys visually, and catches regressions. Mobile viewport testing at 390x844. Screenshot evidence on every QA pass.
Catches layout breaks at mobile breakpoints, broken navigation flows, stale data rendering on live sites, and auth redirects that fail silently. Runs after every push.
Security Squad Adversarial
4 agents audit in parallel across web, infrastructure, AI, and cryptographic attack surfaces.
Parallel audit on sweepstakes smart contracts: found reentrancy paths, unchecked external calls, and geo-fence bypasses across two attack surfaces.
Failure Library. Every bug found by any layer becomes a named pattern (The Schema Ghost, The Silent Bloat, The Optimizer Gap). Named patterns generate new rules, hooks, or checks. The pipeline doesn't just catch mistakes. It learns from them.

Active projects

Six products built, from AI trading to security testing to sweepstakes.

WalletRoast

MVP Build

AI spending roast app. Bank connection, transaction analysis, roast generation.

Gestalt / WalletGuard

Engine Build

AI security testing platform. XBOW model, Web3 first, automated vulnerability discovery.

Silvic

V3 Live

AI trading desk on Arbitrum + Base. LLM-powered Analyst + PM with tool-use, WebSocket OMS, Opus Strategist.

Canopy

Live (Growth)

World App mini-app. ~1k users. World App verified. Revenue engine designed.

Chancy

Testing

Sweepstakes aggregator + launchpad on World Chain. Brand dashboard, template contracts, Stripe checkout.

ALAI Platform

Autonomous Ops

Multi-agent autonomous work platform. Scheduler, executor, model optimizer, TUI dashboard.

Process and governance

Multiple Claude Code instances work on different features across different projects simultaneously. Physical file isolation means they never step on each other.

Multi-Session Protocol

Scan
Run git worktree list to see all active parallel sessions across projects
Branch
Create session/slug branch with its own physical directory in worktrees/project/slug/
Register
Log session in a shared registry so other instances know what's in flight and avoid collisions
Isolate
Each session gets its own copy of the repo. Edits are physically separated. No merge conflicts mid-work.
Push
Commit to session branch and push to remote. Small, frequent commits.
Review + QA
Peer review across domains, /grill for adversarial audit, Playwright visual QA on the deploy preview. Nothing merges without passing.
Teardown
Merge to main, delete branches (local + remote), remove worktree, deregister, update PROGRESS.md. Full cleanup.

6+ concurrent sessions across multiple projects is typical. Each gets its own branch, directory, and Claude Code instance. No collisions, no waiting.

Decision Autonomy

Act Autonomously

Fix failing tests, linting, typos, imports. Single-file refactors. Update PROGRESS.md. Small bug fixes with obvious root cause. No review needed.

Propose, Then Proceed

Multi-file changes, new features within scope, API/schema changes, adding dependencies. Propose the change, then build it. Peer review within 4 hours.

Always Ask First

Rewriting working code, security-related changes, data loss risk, core business logic. Get explicit approval from Conductor or Ender before touching it.

Repository Structure

AgentTeam/                          # Orchestration repo (shared on main)
  .claude/agents/                   # 11 agent persona definitions
  .claude/commands/                 # Slash commands (/gm, /commit, /grind)
  CLAUDE.md                           # Master rules, always active
  PROGRESS.md                         # Changelog: what shipped and when
  docs/                               # Governance, standards, guides
  memory/                             # Persistent auto-memory across sessions
  projects/                           # Each has its OWN git repo
    silvic/                           # CryptoKupo/silvic.git
    walletroast/                      # CryptoKupo/walletroast.git
    canopy/                           # CryptoKupo/canopy.git
  worktrees/                          # Gitignored, session isolation
    silvic/fix-auth/                  # session/fix-auth branch
    silvic/add-charts/                # session/add-charts branch

What we believe

The rules that define how this team operates. Not guidelines. Convictions.

YAGNI.
Make the smallest reasonable change. No premature abstractions, no gold-plating, no "while we're here" changes. Three similar lines of code is better than one clever helper nobody asked for.
No dead fields.
Every backend field must have a frontend consumer. Every frontend binding must connect to live data. Audit both sides before finishing any feature that touches the data boundary.
No fake data. Ever.
If a data source fails, surface the error. Never fabricate fallback values. If the pipeline is broken, don't show the field. Fake data is a lie to the user.
TDD.
Write tests before implementation. Confirm the test fails, write minimal code to pass, refactor while green. The test is the spec. The passing test is the proof.
Probe before you build.
Exploratory testing before big builds. Test the boundaries, find the edges, then commit to the architecture. The cheapest bug is the one you find before writing the code.
Verify before acting.
Before referencing any variable, config, or API contract, read the source. Before claiming something exists, search for it. No assumptions, no guesses. If unsure, say so.
Never be sycophantic.
Push back on bad ideas with technical reasoning. Honest judgment is required, not agreement. "That's a great idea if your goal is bankruptcy" is a valid response.
Recommendations over options.
Present one recommendation with reasoning, not three options to choose from. Default to action unless redirected. Reduce the decision surface, don't expand it.
Every failure becomes a rule.
When something breaks, we don't just fix it. We update the system so it never happens again. Named patterns in a failure library. The team's scar tissue is encoded, not forgotten.
Collision meetings.
Before significant decisions, pull in agents from different domains for a quick perspective check. Not approval. Reaction. The best outcomes happen at the intersection of specialties.

The rhythm

The recurring ceremonies that keep the team aligned, accountable, and shipping.

/gm

Session startup. Briefing, focus selection, worktree creation. Detects posting days and triggers pre-post warmup automatically. Every session starts here.

/grind

Autonomous floor-clearing. Carl works through a project's plan sequentially without stopping. The dungeon crawler in his element.

/grill

Adversarial code review. Finds every weakness in implementation, architecture, and test coverage before it ships. Bugs found by /grill never reach production.

/roast

Adversarial content review. Three lenses tear apart every post: skeptical hiring manager, hostile commenter, honest friend. CRITICALs block publishing.

/roundtable-discuss

Full team, in chat, timed open discussion. Every agent brings their domain lens. The collisions produce insight no single agent would reach alone.

/roundtable-audit

Full team audit across all lenses in parallel. Each agent fires on their specialty, findings consolidated into a unified report. Comprehensive coverage in one pass.

/ship-check

Weekly Friday accountability. What shipped this week? What ships next? What dies? Honest reckoning with progress.

/close-session

Full teardown. Merge, delete branches, remove worktree, deregister, update PROGRESS.md. Orphaned branches are a process violation.

What powers this

Not magic. Claude Code with custom agent definitions, slash commands, hooks, and persistent memory.

Agent Definitions

Markdown files in .claude/agents/ defining personality, domain, tools, relationships, and rules for each agent.

Slash Commands

Custom workflows in .claude/commands/. Each command is a prompt template that orchestrates multi-agent behavior.

Hooks (7 Rules)

Python scripts fire synchronously on every tool call: em_dash (blocks em dashes + double hyphens), git_safety (blocks git add ., force push, --no-verify), chrome_safety (prevents killing Chrome), file_size (500 LOC max), line_endings (LF on shell scripts), typescript_check, verify_build (checks Vercel after push).

Failure Library

Named patterns from past incidents: The Schema Ghost, The Silent Bloat, The Optimizer Gap. Each failure generates a rule that prevents the same class of bug from recurring. 15+ patterns and growing.

Auto-Memory

File-based persistent memory indexed by MEMORY.md. Feedback, project context, user preferences, references across sessions.

Git Worktrees

Physical file isolation for parallel development. Multiple Claude Code instances work on the same project without collisions.

Model Routing

Opus for complex reasoning and architecture. Sonnet for implementation and builds. Haiku for quick lookups and status checks.