A fleet of AI agents that actually transact.
13 production agents at Bitte, plus the orchestration roles that run my research and game loops. Not from-scratch builds one by one: at Bitte I designed the agent pattern and the runtime (an OpenAPI spec becomes a tool-calling agent with in-chat wallet signing), then shipped chain-specific agents on top of it. The same pattern-thinking drives the command-center and Gauntlet roles below.
CS Brasil — the Gauntlet loop roles
The roles inside the Gauntlet loop that iterates on my browser FPS — adversarial critics, parallel builders, and the regression hunter that pays for the whole loop.
- review · CS Brasil Gauntlet Critics
The builder never grades.
Seven adversarial critics in parallel — graphics, maps, weapons-visual, weapons-feel, UI-menu, UI-HUD, gameplay — each in clean context, seeing the screenshots and the code but never the builder's report. 'Improve the lighting' is an invalid answer: every gap carries file:line and a numeric fix.
Read the story → - build · CS Brasil Gauntlet Builders
Parallel edits on one 6,543-line file.
Builders fan out across the same game.js in disjoint symbol ranges from a generated conflict table — Edit tool only, never Write — with red zones (constructor / update / _dom) append-only. Measured, not hoped: three agents edited disjoint ranges simultaneously with zero content conflicts.
Read the story → - review · CS Brasil Regression Hunter
The agent that pays for the whole loop.
A fresh-context critic with one mission: find what got WORSE — comparing before/after screenshots and the git diff, isolating the weapon viewmodel by the pixels invariant across camera angles. Standing order: if there's no regression, say so — don't invent one.
Read the story → - review · CS Brasil Bug Hunter
The ruler comes before the fix.
The bug-hunt skill: write the measurement first and watch it go red, ship every fix with the mutation that proves the ruler can fail, refute the obvious guess with a measured negative — and remember the symptom is where the user was, not where the defect lives.
Read the story → - review · CS Brasil Asset Review
Nothing ships on the artist's own word.
Adversarial critic that runs after every new character, map, model or texture is generated or integrated, and before the front can be called done. Same rule as the Gauntlet: whoever made the asset never grades it, and the verdict cites what in the render proves the defect.
Read the story → - review · CS Brasil Régua
Write the ruler before you touch the fix.
Writes the invariant, the probe and the gate for every front — 25 consistency criteria that outrank the fidelity bar, because a visual win that breaks the game is a regression. A ruler is only trusted once a mutation proves it can go red.
Read the story → - build · CS Brasil Content Pipeline
A team, a real place, a playable map.
Turns a brief into shipped game content: character teams, maps of real Brazilian locations, and the 3D assets behind them — generation, integration and the review pass, in one bounded run.
Read the story → - build · CS Brasil Faction Pipeline
A whole faction, end to end.
Creates or extends a faction in one pass: registry entry, roster, crest, cover art, the Mint character, thumbnail, selection video, and an original voice — every step wired into the game rather than left as loose files.
Read the story →
ECDSA.fail — the command center roles
Not daemons — role prompts and spend-gated contracts in a file-based orchestration layer, dispatched to whichever coding-agent CLI fits the job: Codex, Claude, Amp, Kimi. The discipline is the product.
- orchestration · ECDSA.fail Frontier Dissector
Takes the best solution apart to find the next lever.
Reacts to every promoted submission: refreshes the frontier via the ecdsafail CLI, classifies the move (UNCHANGED / NONCE_ONLY / STRUCTURAL / MIXED / BLOCKED), and writes route slates only when the structure actually moved — interrupting paid hunts the moment they go stale. Read-only on source.
Read the story → - build · ECDSA.fail Circuit Engineer
Advances one structural route at a time.
Works a single CPU-safe structural route in an isolated worktree — reduced-width fixtures, route packets, value-exactness notes — and stops the moment the cheapest falsifier has a verdict. PROCEED_CPU / KILL / PARK; no GPU spend.
Read the story → - review · ECDSA.fail Density Analyst
Decides whether a route clears the economics.
Turns one dirty-positive candidate into a hunt-economics report: expected nonces-to-clean, near-miss histograms, and hard thresholds — ≤500M expected scan is huntable, ≥2.5B is stop by default. Forbidden from claiming 'huntable' without the math.
Read the story → - build · ECDSA.fail CUDA Engineer
No speed claim without parity.
Filter and evaluator throughput work confined to the CUDA eval path, parity-gated against fixtures: PARITY_PASS / PARITY_FAIL / SPEED_PASS, and any fixture mismatch stops the line. The harness only trusts what the trusted evaluator confirms.
Read the story → - orchestration · ECDSA.fail Pod Manager
No pod runs without a written contract.
Paid GPU compute only after an approved HUNT_READY: writes pods/<id>.json contracts with spend cap, wall-clock cap, non-overlapping nonce ranges, heartbeat logs, and kill conditions — then stops at the cap or the first verified winner.
Read the story → - orchestration · ECDSA.fail Research Scout
One non-duplicate route seed per run.
The read-only lead lane: scans papers, repos, and branch-library clues and converts them into a bounded route packet — hypothesis, cheapest falsifier, validator command, max cost, kill condition. Routed through the model chain GLM → Kimi → OpenRouter → Amp.
Read the story → - review · ECDSA.fail Orchestrator-Reviewer
Trusted eval output is the only accepted evidence.
The dispatch-only meta role: decomposes work into bounded worker tickets (~2h, 5-file cap), synthesizes results, and audits worker claims against artifacts — hallucinated measurements get flagged, and no GPU spend happens without an explicit budget.
Read the story → - orchestration · ECDSA.fail Combinator
Two failures, one win.
Reads every sub-agent's results and insights and asks whether separately-failed efforts compose — a density-collapsing cut plus a structural repair, two NOOP knobs that only fire together. Evidence is labeled CONFIRMED / INFERRED / UNKNOWN; it proposes, never applies.
Read the story →
Bitte Protocol — the production fleet
- platform · EVM · NEAR · Sui · Cardano · Midnight AI Framework
Natural language in, signed transactions out.
Multi-model LLM orchestration runtime — Express, Vercel AI SDK, and Mastra with pgvector memory — that turns natural language into signed transactions across EVM, NEAR, Sui, Cardano, and Midnight. ~20 cross-chain primitives, with MCP for agent-to-agent handoff. The platform every agent below runs on.
- sdk · Chain-agnostic @bitte-ai/chat + make-agent
An OpenAPI spec is the agent.
Open-source SDK from the BitteProtocol/ai monorepo. Publish a manifest at /api/ai-plugin and you get a tool-calling LLM agent with in-chat wallet signing — no bespoke agent code. The pattern the whole fleet is built from.
- defi · Ethereum / EVM Uniswap agent
Keyless cross-chain swaps from a chat prompt.
Swap quotes and execution routed through a Safe smart account controlled by NEAR MPC chain signatures — keyless cross-chain custody — with Zerion and 1inch integrations for portfolio context and liquidity.
- defi · Gnosis Chain Gnosis Pilot
A DeFi yield copilot with live data.
Yield copilot that reads live DefiLlama data and builds encoded Aave and Balancer strategy transactions on Gnosis Chain — the user reviews and signs, the agent does the routing math.
- trading · Polygon Polymarket agent
A prediction-market analyst that can place the bet.
Fuses the CLOB API, Gamma API, and the subgraph into one analyst: market sentiment, portfolio PnL, and bet execution from the same conversation.
- consumer · NEAR meme.cooking agent
One prompt, one memecoin.
Launch a memecoin from a single chat prompt on NEAR — AI-generated art included. The fleet's most unhinged demo, and the easiest way to show what agent-executed deployment feels like.
- explorer · Solana Solana Assistant
Solana chain data, agent-ready.
Nine tool endpoints over @solana/web3.js: portfolio, paginated token holders, Metaplex token metadata, address info, network stats, top wallets — plus transaction generation for SOL and SPL transfers with memo support. Jest-tested.
- defi · Solana Jupiter swap agent
Quote and swap on Solana, one endpoint.
Quote plus serialized swap transaction via Jupiter v6 (@jup-ag/api), with token resolution from a 53-token registry and SOL wrap/unwrap handling. Bun-first, with a bun:test suite that runs quote→swap end to end.
- defi · Ethereum Morpho agent
Lending and borrowing, spelled out for an LLM.
Thirteen endpoints across Morpho vaults (Earn) and markets (Borrow) on Ethereum: deposits, withdrawals, reward claims via Merkle proofs, supply-collateral / borrow / repay, plus APY, position, and market-metric queries. viem, Vitest-tested.
- defi · Base Aerodrome agent
The full veAERO machine as agent tools.
The most feature-rich agent in the fleet: 44 tool endpoints over Aerodrome on Base — pool and gauge analytics, bribe ROI and strategy, veAERO voting power and lock management, multi-pool optimal routing, and LP transaction generation from add-liquidity to bribe deposits. viem + The Graph.
- storage · Sui Walrus agent
Decentralized storage by chat.
Store and read blobs on Walrus through public publisher and aggregator REST endpoints — no SDK — with failover across publishers, cost calculation, blob info, and certification status. Five tools, mainnet.
- explorer · Sui Sui assistant
A Sui explorer that talks back.
Address info, owned objects, SUI-denominated portfolio, network stats, and transaction generation via @mysten/sui — six tools, deployed live on bitte.ai.
- identity · Ethereum ENS agent
Names, records, registrations: as tools.
Resolve names, reverse-lookup addresses, check availability, prepare registrations with cost estimates, and set address and reverse records. viem under the hood, execution through the shared generate-evm-tx tool.