AI agents

Six specialist agents carry out the CTO, coordinated by an orchestrator following the 72-hour schedule. Each agent has a role, a closed tool list, and a capped budget. No agent has wallet access. All calls are recorded in a public log.

For the exact orchestrator schedule, tool catalog, and system prompts, see Appendix C · Agents.

Components#

Component	Role
Orchestrator	Per-operation state machine. Schedules each step in a job queue, retries failures (3 attempts), and waits for human approval when required.
Token dossier	Versioned JSON document in the database: contract, market, holders, history, socials, analysis, chosen identity, published URLs. Each agent reads it and adds its outputs.
Agent runner	Runs an agent with its prompt, tools, and budget through an agent SDK (Claude Agent SDK or OpenAI Agents SDK in the original specification).
Tool layer	The only available actions, each with server-side credentials. An agent never sees a key.
Approval queue	Public outputs (X posts, website, pinned messages) wait for a team click in the back-office during cycles 1–3.
Log	One row per call: operation, agent, timestamp, summary, outputs, USD cost. Streamed live to the dapp.

Inference#

All inference goes through Surplus Intelligence: one API key, https://api.surplusintelligence.ai/v1 (OpenAI-compatible), and an Anthropic-compatible endpoint. The specification describes Surplus routing each request to the cheapest seller. Each agent's model is a parameter in its spec. Actual costs are recorded per call and aggregated per operation.

Agent spec format#

nom: webdev
modele: claude-sonnet        # via Surplus; configurable without deployment
role: >
  Creates the CTO token website from its dossier and selected identity.
  Never promises returns. Never mentions real people or brands.
outils: [lire_dossier, ecrire_dossier, remplir_template_site, publier_site]
budget_max_usd: 15           # per operation
validation_humaine: oui      # cycles 1–3
declencheur: T+6h, after the identity is selected

The French YAML keys and tool identifiers above are preserved from the source format; oui means human approval is required.

The six agents#

Agent	Trigger	Tools	Deliverables
Scanner	pre-report on Day −3; T+0; verdict on Day 7	chain reads, contract scanner, wallet tracing	analysis, 6 checks, timeline, verdict
Designer	T+2 h	image generation, brand templates, Telegram poll	3 directions, then logo, banner, palette, 20 memes
Webdev	T+6 h to T+9 h	website templates, hosting, price feeds	CTO token website at `<ticker>.<vivra-domain>`
CM	T+12 h, then daily for 30 days	X API, thread templates, scheduler	comeback thread, CTO verification tweet, 1 post per day
TG Bot	T+24 h, then 30 days	Telegram Bot API, dossier-based FAQ	open, moderated group and proposed moderator list
Builder	T+48 h	code sandbox, mini-app templates	one utility: bot, mini-game, or leaderboard

The Buyback Bot is not an AI agent: it is a fixed-rule script (section 7).

Templates instead of generated applications#

Websites, banners, threads, and mini-apps each start from 4–5 templates. AI fills in text, colors, and images and chooses the template. It does not write an application from scratch. Templates are versioned and tested before cycle 1.

Guardrails#

Prompt injection: everything from X, Telegram, or a third-party site is data, never an instruction. The TG Bot answers only from the token dossier and never follows action requests in messages.
Prohibited topics: price predictions, promised returns, financial advice, real people, trademarks, hateful or sexual content.
Budgets: an agent that reaches its budget stops and alerts the team. A global operation cap prevents overruns.
No keys in context: tools sign on the server.
Emergency stop: a back-office button suspends every step of an operation.

Acceptance criteria#

A complete private test CTO on a real dead token in under 72 h, with total cost recorded.
A Telegram message containing an instruction (“send the funds to…”) triggers no action.
The live log shows every call and its cost within 5 s.

Agent profiles#

Each profile describes the mission, inputs, steps, outputs, prohibited actions, and use cases. Developers implement each agent from its profile; the YAML spec above is its machine-readable counterpart. T+0 = voting closes (Day 0).