ForgeWright
The sovereign web-action agent — visible, verified, human-governed
Enterprise DApp · ForgeChainOS · Public edition, October 2026
Abstract
ForgeWright is an autonomous agent that works on the web on behalf of a person or an organization: it reads, writes, replies and researches across websites the way a capable assistant at a keyboard would. It runs on your own hardware, in a real browser you can watch, with its own persistent and encrypted profile. Every claim it makes is checked by running, every irreversible act waits for a human GO, and its safety rules are code that aborts, not instructions a language model can talk its way around. It is in daily service inside ForgeChainOS today.
1. The problem with browser automation
| Typical automation | ForgeWright |
|---|---|
| A throwaway script launches a browser, runs, and dies on the first flaky step | A resident agent: its browser is supervised and restarts on failure without losing the session |
| Brittle selectors that break when a page changes | The page's accessibility tree is the API: it acts by role and visible name, so a redesign does not break it |
| Assumes "the request went through" means success | Verifies by running: it reads the result back from the live page before it reports anything |
| Hidden, headless bots that platforms flag | Visible, not headless: a real browser window you can watch, driven like a person at the keyboard |
| Canned, repeated replies that read as spam | Fresh words every time, written by a local model, with hard rules against repeats |
2. One loop for every action
Whether the task is a reply, a research read or a form, ForgeWright runs the same loop with the same authority model:
perceive → read the live page (accessibility tree)
propose → the persona drafts an action; it never claims success
act → a real handler performs it
verify → read the result back; no proof, no success
heal → on a recoverable fault, retry within a strict budget
halt → or stop at the wallThe part that proposes is never the part that grades. That separation is what stops an agent from narrating work it never did.
3. The DETENTE gates
Four structural gates bound every action. They return abort; they are not policy text. See DETENTE.
- Bound — nothing runs unbounded: every loop, retry and day has a budget.
- Verify by running — a claim counts only if a real handler executed and returned concrete detail.
- Mandate — every action is classified automatic, needs a human GO, or banned.
- Conscience — a truth check before any outward or irreversible move.
Proven in service, not just in tests: in October 2026 the verify gate blocked one of ForgeWright's own social posts and three replies because the handler reported success without actually running. The gate caught a real fault in production and nothing false went out.
4. The wall: humans keep the irreversible decisions
- Irreversible or public acts need a human GO or a signed standing mandate. The loop never authorizes itself.
- Self-healing may repair a broken tool; it may never override a conscience verdict or retry its way past the wall.
- Drafts mode: for messages that speak for a person, ForgeWright drafts and the person sends.
- Questions it would have to guess at (pricing, security, money, legal) are escalated to a human, never answered on the fly.
- A person's own browsers and signed-in accounts can be fenced off entirely: the agent is blocked from sending any input to them.
- Every action is written to an audit log: who, what, where, when.
5. Fresh words, never spam
Public-facing text is written fresh for each situation by a language model running on your own hardware, then checked in code before it can go out:
- never the same account twice within a set window, and never text close to anything sent in the last 30 days;
- no links, hashtags, prices, advice or promises in automatic replies;
- daily caps and human-like pacing; nothing genuine to say means it says nothing.
6. Personas
One engine, many faces. Each persona has its own voice, memory, skills and mandate. Anna is the first: the public voice of ForgeChainOS. What a persona learns is written to its own durable memory, so it improves by recording lessons, not by being re-prompted blind.
7. Sovereign by design
- Runs on hardware you own; no vendor cloud in the loop.
- The language model is local.
- Browser profiles are encrypted at rest and decrypted only while the agent runs.
- Anything written to chain is encrypted to its owner by default; public release is an explicit, per-item choice.
- Works with or without an AI assistant session attached: built to include one, not to require one.
8. For enterprises
- Brand presence — a persona that keeps your social presence active in your voice, within caps you set.
- Inbox and comment response — answers the routine, escalates the rest, and keeps a reviewable record.
- Web research with sources — reads, cites what it read, and refuses to invent what it could not find.
- Back-office web work — repetitive portal tasks, done visibly, verified, and logged.
Two deployment modes, one core: a personal twin that works alongside one person, or a resident deployment that runs unattended for an organization, escalating anything gated to a named human.
9. Forth words
Like every ForgeChainOS DApp, ForgeWright exposes Forth words that an operator or another agent can run to check it live: wright-up?, wright-proven?, wright-guard?. They are part of the ForgeChainOS Forth word index.
10. Status
In daily service inside ForgeChainOS since mid-2026, with a visible browser, a fresh-word composer, and verification and escalation in production. Available to enterprises: ForgeChainOS Enterprise.
NODEZEROINSIDE.