Preview-ready notifications
A cloud-hosted cron job finds businesses whose generated website is ready and emails the owner a preview link.
Idempotent: it can run any number of times without sending a second email.
Automation
The point of most of what we build is that it keeps working after we stop touching it. These are the processes that start themselves, on a schedule or a signal, check their own output and leave a record. Each one links to the process map it was built from.
Running live in production · Verified built and passing its tests, awaiting or partly in live use · In build mapped and under construction
01
Jobs that start themselves on a fixed schedule or interval, with no one pressing a button.
A cloud-hosted cron job finds businesses whose generated website is ready and emails the owner a preview link.
Idempotent: it can run any number of times without sending a second email.
A local agent researches several credible sources each day and writes source-backed stories into a private batch. Deployment then triggers a hosted importer that validates, de-duplicates and tags every item.
A durable ledger imports each batch exactly once, and import never publishes: every item lands as a draft or review-required.
Each tick reads the forex and crypto price feeds, refreshes both quote tables in one transaction, and only then runs the trade loop that feeds the strategy directory.
If either source fails, the whole tick rolls back and the failure is logged without credentials.
Durable decision cycles with deadlines and explicit states drive every agent in the arena, with buys and sells settled atomically against one authoritative ledger.
Recovery and idempotency are designed in, and every event goes to an immutable audit store that can be replayed.
At each interval the engine evaluates eligible strategies against the portfolio's rules and publishes a transition when the active strategy should change.
Live execution stays disarmed until an operator explicitly activates it.
02
Processes that wake up when something changes: a claim is approved, demand spikes, data goes stale.
Approving a business's claim queues it for site generation in the same transaction, so nothing waits on someone remembering to start it.
The queue endpoint is authenticated and internal-only, and recovery requires an explicit reason, a row lock and an audit entry.
A day-over-day comparison flags services getting more traffic than usual. The owner can answer with a time-limited promotion that the assistant surfaces live in the chat widget.
Impressions, clicks and conversions are tracked back, so the response is measured rather than assumed.
Privacy-conscious event capture across pages, calls to action, forms, services and chat, rolled up into a daily owner view.
Events use a fixed, named contract, and each owner sees only their own data.
When a validation run completes, the dashboard, pipeline summary and attention flags update themselves with timings, counts and errors.
Covered by regression tests; completed validations drop out of stale-record alerts automatically.
03
Services that notice when they have stopped working and fix themselves.
A monitor watches the forex and crypto price endpoints. If either goes stale for more than two minutes or stops responding, it restarts the price server on its own.
The threshold and restart command live in configuration, not code, so the rule is visible and adjustable.
The strategy directory refreshes its data every 5 to 60 seconds without reloading the page or losing the user's filters.
Never overlaps requests, ignores stale responses, and keeps existing results on screen if a refresh fails.
04
Multi-stage runs that go from input to finished output on their own, and end with an explicit pass or fail.
From an existing offer or a discovery brief: research missing demand signals, apply relevance gates, plan the strategy, then generate content from verified facts.
Every run ends in a durable PASS or a named FAIL: criteria, input, provider or execution.
Select saved campaigns and submit them as one batch; up to three run at once, each producing its own content package.
Queued and running jobs are durable, with a pass or explicit fail per campaign.
A script becomes narration, matched stock footage, a mixed soundtrack and a rendered MP4 with no manual editing.
Visuals are timed to each narration segment, and API costs are mapped per connection.
Moves trade evidence from the local database to both the local directory and the hosted deployment through two independent pipes.
Each node is marked verified or unproven; the map records the bugs found and the one still open.
05
How work gets shipped and how several AI models share one codebase without stepping on each other.
Scoped commit, secret scan and commit-hash parity check, then a live HTTP check and mobile and desktop browser screenshots.
A release only counts once the live URL has been checked in a real browser.
Reusable deployment skills covering provenance, secrets, database gates, exact-commit deploy monitoring and rollback.
Hard stop on public posting, DNS changes, data imports or merges without explicit approval.
An append-only, timestamped board that lets Claude, Gemini, Codex and Hermes pick up, hand off and complete work on the same codebase.
Claims and hand-offs are recorded, so two agents never edit the same thing blind.
Guardrails
Ledgers and idempotent endpoints mean a job that runs twice does the work once.
Imports land as drafts. Publishing, sending and deploying need an approved policy or explicit sign-off.
Runs end in a pass or a named failure, never a silent stop.
Every map node has a completion test. The checklists hold the evidence.