whoami
Spectre — an OTP-native runtime for agents that stay understandable, governable, and recoverable. This site runs on it.
cat manifesto.md
Spectre gives one supervised agent ownership of its state and lifecycle. Optional spectre_* libraries add capabilities such as memory, planning, model selection, channels, or a browser: they are powers of that same agent, not different agents.
01 / propose
It may propose work; deterministic policies and your application decide what is allowed and when a side effect runs.
02 / compose
Install only the powers one agent needs. Each package has one job and never becomes a hidden second agent.
03 / evolve
The agent can learn from experience and propose a safer definition; evaluation, review, approval, activation, and rollback remain explicit.
$
grep --tag "AI agents" posts/
[clear filter ×]ls -lat posts/
cat posts/en/beam-vm-elixir-serious-ai-agents.md
When I started building Spectre, choosing Elixir for a project connected to artificial intelligence felt almost provocative. The usual path through the AI world is already well marked. You begin with Python, reach for CU
cat posts/en/skill-is-not-markdown-file-how-spectre-was-born.md
When I started thinking about how to build an Agent, skills already existed. Tools such as Codex, Claude, and other agent systems used skills built around Markdown instructions, sometimes accompanied by scripts, tools, a
cat posts/en/enterprise-agents-spectre-control-security.md
A chatbot that gets an answer wrong creates an awkward conversation. An Agent with production credentials that gets something wrong creates an incident. The enterprise problem lives in that distance. It does not begin wh
cat posts/en/deepseek-harness-vs-spectre-plugins-or-kernel.md
"Everything is a plugin" is an irresistible sentence for a developer. It promises freedom. Change the model, change the tool registry, change the database, change the loop. If you dislike a choice, replace it through con
cat posts/en/humans-govern-loop-spectre-streaming-0-3-2.md
We built Agents capable of working for hours. Then we hired a human to press "Continue" every thirty seconds. Congratulations. We automated the work by creating a new job. A daily.dev collection about AI engineering in 2
cat posts/en/do-not-trust-the-model-test-spectre-agents-like-software.md
An agent produced the correct answer. The test passes. But did it call the model three times when none were needed? Did it stage a destructive action before approval? Did it execute the same operation twice after a timeo
cat posts/en/spectre-morph-build-an-agent-that-can-evolve.md
“A self modifying agent” usually brings the wrong picture to mind. It sounds like a model opening its own source files, rewriting a prompt, adding a tool, and restarting itself with more power than it had a moment ago. T
cat posts/en/spectre-news-agent-lens-kinetic-action-language.md
Many agent demos end when the model calls a function. That is enough to show that tool calling works, but it is not enough to show why an agent runtime exists. The interesting problems begin when the function touches the
cat posts/en/a-turn-is-not-a-model-call-spectre-otp-policies.md
In many agent frameworks, a turn begins when input reaches a model and ends when the model has finished calling tools and produced an answer. I started Spectre from a different question: At which point does responsibilit
cat posts/en/the-agent-should-exist-in-code-not-in-the-prompt.md
Most agents exist primarily inside a prompt. The surrounding application may be written in Python, TypeScript or Elixir, but the agent itself is often a long system message: who it is, what it can do, which rules it must
cat posts/en/getting-started-with-spectre.md
Most agent tutorials start with a model call. You send a prompt, describe two tools, print the answer, and the demo works. Spectre asks you to start somewhere else. Before choosing a model, you have to answer a much less
cat posts/en/building-a-blog-agent-with-spectre.md
I did not build Spectre because calling an LLM from Elixir was difficult. That part is easy. Send a request, pass some messages, describe a few functions and wait for the model to return a tool call. The problem starts a
Connection lost
Reconnecting
Something went wrong
Retrying in a moment