Loops
Loops in 📝MythOS are repeatable agent processes written as memos — what starts one, which memos it must load, and the steps it runs — and they are an upcoming extension of 📝Augmentation, the identity layer your AI already carries into every session. Every agent run drags a payload of context nobody can see, so the same process costs a different amount each time it fires and no one can say afterwards which memos actually arrived; a loop makes that payload a declared contract, classed by what guarantees each load and measured in tokens and characters before anything runs. It is built for 📝Oracle creators who run the same process often enough that its context and its cost are worth pinning down.
Key Capabilities
- The definition is a memo — no second record, so an agent amends a loop through the same object it reads
- A declared load contract — every memo a run receives, classed as hook, gate, trigger, or search
- Weight before the run — tokens and characters per memo, per class, per loop, and across a whole portfolio
- A canvas and a memo view — one definition read either as a step graph or as the document it actually is
- Instruction drafting on the canvas — arrange notes, routers and filters beside the loop's steps, connect them into a labelled flow, and edit instructions in a drawer; draft saves stay separate from executable steps
- A portfolio dashboard — state, mechanism counts and context weight per loop, filed into folders and pinned
- An execution engine behind a human gate — a run drafts with the declared model or on your own device, suspends at the loop's declared approval gate in the review queue, and releases only on a recorded decision
- Runs that start themselves — a loop can run on a schedule or when something arrives from a connected source such as a webhook, RSS feed, newsletter or tagged memo; each arrival is counted once
- Channels, Devices and Sessions — channels collect a loop's messages, asks and alerts, devices are the machines you register to run loops, and each device run keeps a session you can follow and answer
- A run record for every pass — status, per-step spend, duration, and where the output landed, rendered as History beside the portfolio
- Spend that rides your budget — every model call a run makes is metered into the same monthly caps and alerts as chat
Getting Started
Loops is in build and not yet live on mythos.one. When it opens it will be an Oracle-tier surface; a loop is created from a starting template and never fires the moment it exists, so a definition can be read and weighed before it is armed.
FAQ
- What is the difference between a loop and a run? A loop is the definition, an ordinary memo whose trigger, load contract and steps are sections in its body. Each time it fires is one run of that loop, with its own record.
- What is a load contract? The set of memos a run receives, each classed by what guarantees it: a hook is injected every run, a gate must resolve or the run aborts, a trigger fires on a named condition, and a search retrieves at run time. Only hooks and gates are certain to land, so a loop reports two weights: what always arrives before step one, and a ceiling that adds every conditional or retrieved load.
- Where does a loop actually run? Its Runner setting chooses. On the engine, MythOS runs it on its own servers, each step one model call with no tools, metered into your plan's AI budget. On an external runner, one of your 📝Devices claims the run and works it as a full Claude Code session with tools, on your machine and your own subscription, then reports outputs and token use to the run record.
- Do loops run themselves yet? The engine is built: a queued item becomes one run that drafts with the declared model, suspends at the declared approval gate, and dispatches on approval, leaving a run record either way. Nothing fires for customers until the surface opens; for now it is limited to MythOS staff.
- What keeps a run from publishing on its own? A loop that declares a hold gate cannot even compile without a step for that gate to suspend, so enforcement is at build time, not by promise. A suspended run waits in the same review queue as every other held draft until its owner decides.
Related
- 📝Implicit Loading — the sibling mechanism that guarantees your identity context loads before an agent acts
- 📝Semantic Search — the retrieval behind a search-class entry in a load contract
