IYInnovative You

Build record

Amúyẹ

built

An agent procurement system that commissions specialist work for protocol assessments, learns from execution, and applies that experience to later jobs.

Classification
Software / Agent system / Hackathon project
Last updated
Source
Open repository
Live
Open site
Amúyẹ execution console showing a completed controlled protocol assessment, specialist purchases and the Sibyl memory comparison

What

Amúyẹ accepts a protocol-assessment objective, budget, deadline, priority and hard constraints, then purchases useful specialist evidence progressively and can stop before unnecessary deeper work.

System facts

  • Protocol assessments are decomposed into specialist capabilities.
  • The execution controller enforces dependencies, gates, budget and early stopping.
  • Sibyl stores execution history and operational lessons across sessions.
  • In a controlled repeat run, recalled experience changed the purchase from three specialists costing 95 to two costing 35.
  • Recalled experience can change the plan, but it cannot override the job's hard constraints.
  • Risk synthesis can use a real Virtuals ACP provider purchase, with ACP spend counted against the client budget.
  • The browser console supports proof modes and new assessments, while the programmatic API exposes the same constrained assessment path.

Boundary

  • Amúyẹ is limited to evidence-only, risk and security protocol assessments; it is not a general-purpose agent.
  • Live ACP spending requires explicit confirmation, and not every specialist is purchased through ACP.

Why

Protocol assessments often need several kinds of evidence, but not every job needs every specialist. Amúyẹ uses execution history to adjust later procurement while keeping the current job's constraints authoritative.

Evidence

Log

  1. Published the Amúyẹ browser console.

    Configured the static frontend for the verified public deployment without widening its live-purchase authority.

  2. Separated public proof from paid ACP authority.

    The public assessment path can sync partner proof while live ACP spending remains behind the controlled execution boundary.

  3. Made assessment history survive restarts.

    Assessment jobs now persist across process restarts, keeping prior runs available to the browser console instead of losing them with runtime memory.

  4. Made failed live purchases stop safely.

    A failed live ACP purchase no longer triggers an automatic paid replacement attempt.

  5. Connected planning to specialist execution.

    Built the controller, specialist stages and ACP path, then connected the execution console to the real application runtime.

  6. Proved memory changed the purchase.

    Under the same controlled task conditions, the cold run purchased viability, risk and security for 95. The memory-backed run recalled prior execution experience, stopped after risk and spent 35.

  7. Settled a real specialist purchase on Base.

    Virtuals ACP job 75660 moved 0.1 USDC into escrow and completed settlement across the provider, evaluator and platform treasury.

  8. Made memory yield to a hard constraint.

    A recalled lesson normally allowed security to be skipped. When security became mandatory, the controller kept the useful ordering but overrode the conflicting memory rule for that job.

  9. Made past execution change a future plan.

    Built the first Sibyl-backed planning proof, where operational experience persisted across sessions and could alter later specialist procurement.

Stack

  • Python
  • TypeScript
  • Sibyl
  • Virtuals ACP
  • SQLite
  • Base