Meet Aielia — an agent that takes on a job and works it through: researching a list, going through your files, drafting and sending the update. A model on its own improvises. Aielia doesn't, because a harness decides what it may do at every step.
The model proposes; the harness decides. Eleven layers govern which actions are allowed right now, what Aielia is allowed to believe, whether an answer holds up, and how much a job may cost — every real turn, in code rather than in a prompt. It's open source, and you can build a harness yourself.
Diagram: a simple agent loop (input → LLM call → tool call → output) compared with a full Build A Harness — an 11-layer architecture across 27 node types that adds caller state, a world model, reasoning, a 5-tier control layer, planning, execution, 9-layer verification, recovery, memory, optional learning, and an output reviewer pass.
Aielia isn't only for one-line questions. Hand her a piece of work and she carries it out step by step, using tools — and at every step the harness, not the model's mood, decides what is allowed to happen next.
> Find the pricing page for each vendor and
summarise the plans:
- Acme
- Globex
- Initech
- Umbrella
- Hooli
budget calibrated on the first two items, per item
Acme found
Globex found
Initech not_found — 3 dead ends in a row, moved on
Umbrella found
Hooli truncated_while_productive — budget reached
reply built from those outcomes; nothing omitted
> Email that summary to the team.
staged — send_email waits for your approval.
The exact action is recorded; nothing else can
be swapped in after you say yes.Each item gets its own bounded search, so one dead end can't eat the budget for the rest. Every item ends found, not found, or cut short while still productive — and the reply says which.
Read and reason over a workspace folder you choose. Off by default, and available in the terminal and desktop apps rather than the browser tab. Writes are staged for your approval.
Compose the message, then stage the send (email is available in the terminal app today). Email, file writes and shell commands never run on the model's say-so; they wait for your approval.
A reminder list for what's pending (notes for now; it doesn't alert you), a plan for a compound request, and memory of what you've told her — with the source of each fact recorded.
file and shell access are opt-in · web search needs your own key · Apache 2.0
Prompts can be ignored; a harness can't. These controls run as code around every turn, so behaviour stays inside the lines whether the model is having a good day or not.
Every tool call is checked against a live control state — normal, cautious, or blocked — before it runs. It tightens after failures, and the answer is a deterministic allow, deny, or require-approval.
What the model merely infers is held to a higher bar: only high-confidence inferences reach lasting memory on their own, and medium ones wait for your confirmation. Each fact keeps its source, contradictions are flagged, and web pages and command output are treated as data the model must never obey.
Replies are checked and reviewed before they reach you, and ones that used tools carry their evidence with a plain status: verified, unverified, contradicted, or no evidence.
Per-item budgets, dead-end detection, a confirmation step before a big batch, and a hard ceiling on tool calls per turn — so a stuck job stops rather than spirals.
Failures are classified and met with a recovery strategy rather than a blind retry. The harness checkpoints mid-turn, so a crash resumes the same job instead of restarting it.
Writing a file, running a command, and sending an email are staged as an exact action and applied only on your approval, unless you deliberately switch that prompt off. A message that reads as consequential is stopped after one quick classification call, before the harness or any tool runs.
11 layers · every real turn · enforced in code, not in the prompt · Apache 2.0
Nothing to install. Open myaielia.com/try, paste a model key in Settings, and start talking. The browser tab is chat-only; file, shell and email tools live in the terminal and desktop apps.
curl -fsSL https://myaielia.com/install.sh | sh installs a single binary — no Node needed. On Windows: irm https://myaielia.com/install.ps1 | iex. Already have Node? npx @buildaharness/aielia works too. First run walks you through picking a model.
The same chat UI in a native window, with a folder picker for file access. Built from source today, and it needs a Rust toolchain.
Aielia is in public alpha. Found something broken? Tell us on GitHub.
Aielia is the reference application for Build A Harness — an open-source visual canvas and runtime for the same control layers. Design your own agent harness, compile it to LangGraph, CrewAI, Mastra, or Microsoft Agent Framework, and read how we measure which features earn their cost.