eAI — Multi-Model Engineering Agent Fleet

A tech-oriented, multi-model agent fleet with an explicit orchestrator/worker hierarchy. Expensive models plan and review; cheaper models execute. Every agent declares the model tier it runs on.

eAI re-cuts the agency-agents collection (MIT) for engineering work and wires it for cost-aware model routing. Where the upstream is a broad agency (marketing, sales, finance, design, …), eAI keeps the engineering, security, testing, and technical-specialist divisions and adds the missing piece for production use: a routing layer that decides which model runs which agent.

It also adds a native blockchain division (workers/blockchain/) for multi-client execution-layer work — Geth-fork core, Erigon/Reth/Besu/Nethermind porting, consensus parity, p2p/sync, RPC parity, fleet monitoring, and node ops — plus a client-port mission playbook where consensus parity is the sacred gate and every hash mismatch is a P0.

The model

   PLAN / DECIDE        👑 Prime  — Claude Fable 5     architecture · irreversible calls · conflict resolution
                        🎯 Lead   — Claude Opus 4.8    decomposition · dev↔QA loop · gates · merges
        ─────────────────────────────────────────────────────────────────────
   EXECUTE              🧠 Senior — Claude Sonnet 4.6  correctness-critical · security · debugging
                        ⚙️  Bulk   — Kimi K2            high-volume codegen · CMS/UI · refactors
                        ⚡ Fast   — Claude Haiku 4.5   lookups · formatting · classification

• Two orchestrators (orchestrators/) run on the strongest models because their mistakes are the most expensive. Prime (Fable 5) owns architecture and irreversible calls on a strict budget (≤5 activations/phase); Lead (Opus 4.8) runs the day-to-day pipeline.
• 68 workers (workers/) carry a model: tier in their frontmatter. The orchestrator routes each task to the cheapest tier that can do it correctly.
• Routing (routing/) is the source of truth: models.yaml maps tiers → model IDs, pricing, and escalation triggers; routing/README.md explains the why.

See the full ROSTER for every agent and its tier.

Why multi-model

Single-model fleets either overpay (everything on Opus) or under-deliver (everything on Haiku). eAI matches capability to need:

Tier	Model	$ / MTok (in/out)	Sweet spot
👑 Prime	claude-fable-5	10 / 50	The few calls that are expensive to undo
🎯 Lead	claude-opus-4-8	5 / 25	Long-horizon orchestration, loops, gates
🧠 Senior	claude-sonnet-4-6	3 / 15	Security & correctness-critical reasoning
⚙️ Bulk	kimi-k2	Moonshot	Cost-efficient high-volume code generation
⚡ Fast	claude-haiku-4-5	1 / 5	Read-only sweeps, formatting, classification

Anthropic per-million-token rates as of 2026-06. Kimi K2 runs on Moonshot AI's API (or any OpenAI-compatible runtime). Adaptive thinking is the default on the Anthropic tiers; tune effort per route. On Fable 5, omit thinking rather than sending disabled (it 400s).

Quick start

1. Tokens via env vars

cp .env.example .env    # fill in ANTHROPIC_API_KEY + MOONSHOT_API_KEY — never commit .env

ANTHROPIC_API_KEY (X-Api-Key) pays for the orchestrators and senior/fast workers; MOONSHOT_API_KEY (bearer) pays for bulk-tier Kimi work. Shell exports override .env. To pin config per project instead, use the "env": {...} block in .claude/settings.json.

2. Install the agents

./scripts/install.sh --tool claude-code        # installs orchestrators + workers to ~/.claude/agents/

3. Launch a lane

One Claude Code process talks to one endpoint, so eAI runs two lanes:

./scripts/claude-eai.sh                  # Anthropic lane — Fable/Opus/Sonnet/Haiku tiers
./scripts/claude-eai.sh --bulk           # Kimi lane — bulk work on real Kimi K2 tokens
./scripts/claude-eai.sh --continue       # resume your latest Anthropic-lane session

On the Anthropic lane, the Bulk tier falls back to claude-sonnet-4-6 (see fallback: in models.yaml). The --bulk lane points Claude Code at Moonshot's Anthropic-compatible endpoint (ANTHROPIC_BASE_URL=https://api.moonshot.ai/anthropic), so those sessions genuinely bill to Kimi. Activate an orchestrator and let it route:

"Activate the Lead Orchestrator. Mission spec: ./SPEC.md. Run phase 1."

Bulk assignments can also run headless on the Kimi lane: ./scripts/claude-eai.sh --bulk -p "implement the dashboard from design/spec.md". The Kimi CLI remains an alternative bulk runtime — see integrations/kimi.

4. Drive it with native orchestration

The installer also drops eAI's saved workflows into ~/.claude/workflows/ and slash commands into ~/.claude/commands/:

/eai-mission ./SPEC.md                                    # Lead Orchestrator: decompose → tier-route → dev↔QA
Run the eai-devqa workflow with args {tasks:[...]}        # deterministic implement→verify→retry(≤3) fan-out
Run the eai-parity-sweep workflow with args               # consensus parity: hash/stateRoot/receiptsRoot sweep
  {canonical:"http://c:8545",candidate:"http://x:8545",from:80000,to:80500}
/loop 10m /eai-parity-watch http://c:8545 http://x:8545   # recurring drift watch (Esc to stop)
/effort ultracode                                          # opt this session into workflow-everything mode

Watch running workflows with /workflows. The decision guide for Agent vs Workflow vs /loop vs ultracode is in coordination/orchestration-runtimes.md.

Regenerate after editing

SRC=/path/to/agency-agents ./scripts/gen-workers.sh   # rebuild tier-tagged workers from upstream
./scripts/build-roster.sh                             # rebuild ROSTER.md from frontmatter

How a mission flows

1. Prime reads the spec, decomposes it into phases, and sets the architecture (ADRs for irreversible calls).
2. Lead takes each phase, breaks it into tasks, and assigns each to a worker tier via routing/models.yaml.
3. Workers execute; testing workers gate every task with evidence, not claims. Lead runs the dev↔QA loop — max 3 retries, then escalate.
4. Escalations climb fast → bulk → senior → lead → prime. Only architectural or irreversible decisions reach Prime.

Every transfer uses an eAI handoff template — context loss is the #1 cause of multi-agent failure.

Layout

orchestrators/   👑 Prime (Fable 5) · 🎯 Lead (Opus 4.8)
workers/         🧠 senior · ⚙️ bulk · ⚡ fast — engineering, blockchain, security, testing, specialized
routing/         models.yaml (source of truth) + routing guide
coordination/    handoff templates + dev↔QA loop + client-port playbook + orchestration runtimes
.claude/         saved workflows (eai-devqa, eai-parity-sweep, eai-rpc-parity) + slash commands (/eai-mission, /eai-parity-watch)
scripts/         gen-workers.sh, build-roster.sh, install.sh
integrations/    claude-code, kimi
ROSTER.md        full agent index (generated)

Credits & license

Derived from msitarzewski/agency-agents under the MIT License. eAI's contributions — the orchestrator/worker hierarchy, the model-routing layer, tier tagging, and the coordination rework — are likewise MIT. See LICENSE.