Build one Elixir agent. Give it only the powers it needs.

An agent is a supervised system, not a prompt or a costume for a model. Spectre gives each agent one canonical owner for state, explicit messages at every boundary, and recovery designed in from the start — the way OTP treats everything else.

cat POLICY.md

Anything protected crosses one boundary:

• - models and routes may propose work
• - protected work must pass a deterministic policy
• - approval changes state but does not execute the work
• - execution happens only through an explicit host call
• - every terminal outcome is returned as data

A Spectre agent should read like a map, not a magic trick.

// The whole agent is one module: the model, the router, a protected action, and the policy guarding it.

defmodule MyApp.SupportAgent do
  use Spectre.Agent

  model(MyApp.LLM, purpose: :smart)
  router(via: [:regex, :embedding, :llm_classifier])

  actions MyApp.SupportActions do
    protect(:delete_account, with: :delete_account_confirmation)
  end

  policy :delete_account_confirmation do
    request(:confirm_delete_account)
    accept(:confirmed_delete, regex: ~r/^yes, delete it$/i)
    reject(:cancel_delete, regex: ~r/^(no|cancel)$/i)
    attempts(3, then: :cancel_pending)
  end

  interrupt :HELP, regex: ~r/^(help|menu)$/i do
    reply(:help)
  end
end

man spectre-morph

How an agent can improve itself

Spectre 0.3.2 can record explicitly enabled, redacted experience, inspect the active agent definition, and propose bounded changes to skills or configuration. The proposal is inert data: it cannot publish, approve, or activate itself.

1. 01 / observe

   Experience

   The host opts in and records bounded evidence with sensitive values redacted.

2. 02 / reflect

   Reflection

   Spectre compares what was declared, what is active, and what was observed.

3. 03 / propose

   Proposal

   Forge may suggest a closed change, but that proposal has no authority of its own.

4. 04 / govern

   Evaluation and activation

   Protected tests, review, approval, activation, and rollback guard every change.

Self-improving does not mean self-authorizing: evaluation, human review when required, activation, and rollback stay outside the model.

curl -s spectre_ecosystem/status.json

Compatibility with the current core

Every satellite library is rebuilt and tested against the commit that Spectre core is on today. The table is the published result of that run, not a promise: this page reads it once a day and shows exactly what came back.

// elchemista.github.io/spectre_ecosystem — the full report, refreshed by its own scheduled run