Premium, enterprise-grade agent builder that brings software engineering discipline to AI agent development. Spec-driven builds, deterministic testing, staged deployments.
docs/spec/README.md — Agentelic product specification
Agentelic is the engineering layer:
SpecPrompt (Standards) → Agentelic (Engineering) → OpenSentience (Runtime) → Graphonomous (Memory)Every agent starts from a SpecPrompt specification, is tested deterministically, and deployed through staged rollouts with compliance gates.
Spec-driven build pipeline (parse → generate → compile → package)
Deterministic testing derived from acceptance criteria
Staged deployment (staging → canary → production) with governance gates
MCP tool surface for AI-assisted agent development
index.htmlindex.html is emitted by node build-site.mjs from records/surface.json + src/. A direct edit is overwritten on the next build. Then node launch-gate.mjs reads the emitted artifact and refuses to publish it if anything in it disagrees with the record.
records/surface.json → build-site.mjs → index.html → launch-gate.mjs
↑ ↓ idanim.js ↓
frozen record recomputes the contact.js 148 checks over the
template hashes, ↓ ARTIFACT, not the
then records the records/build.json source
sha256 of every ↑
input and output the gate recomputes every one of themDo not hand-type a check count. The gate prints its own total; 148 is what it printed on 2026-08-17, up from 103 in a single session.
npm run test:launch — build then gate. **Set the Cloudflare Pages build
command to this, never to a plain build**, or an unproven artifact gets served.
The two template SHA-256 values on the page are not copied from the record.
They are recomputed from priv/templates/ with compute_template_hash/1's own algorithm (lib/agentelic/templates/registry.ex) and the build exits non-zero on a mismatch. That is the site's whole evidence claim.
No `mailto:` anywhere (Travis's call, 2026-08-11) — the gate refuses one.
Corrections go to a Formspree form (https://formspree.io/f/xaewoadr, the endpoint computedriven.com posts to; SHELL.md r9, ruled by Travis 2026-08-17). It is a real <form action method="POST">, so it posts with JavaScript off; /contact.js only upgrades it to an inline reply and prints "sent" only on an actual 2xx. The gate refuses if the action is not the endpoint the record declares, or if the _gotcha honeypot is missing — a honeypot dropped in a refactor fails silently. The issue tracker stays as the second, public route.
**The two gate holes SHELL.md r6 names are closed here, and both were PROVEN
open first.**
The retraction blocklist counts now; it used to exempt by class name.
Appending the retracted sentence three times to retracted.paragraph put "Start building today" on the artifact four times and the gate approved it, 114/0. It now bounds occurrences in both directions: zero outside the retraction, at most one inside it.
Nothing proved the artifact came from this build. A build made to throw
after its own checks left the previous index.html on disk and the gate passed over it. records/build.json now records the sha256 of every input read and every file written, as the last act of a successful emit, and the gate refuses on any drift — plus an independent recompilation of the CSS, /contact.js and every literal run of src/landing.html, so a deleted manifest is not the only thing standing there.
The computed-colour resolver (SHELL.md r7/r8). Contrast checks read
declared tokens; r7's header-CTA defect (.top nav a at 0,2,1 beating .btn at 0,1,0, painting --fg2 on the accent at 1.19:1) passed every one of them on the sibling surfaces. The gate now resolves the cascade over the emitted artifact against each .btn's real ancestor chain, at 1600/1280/800/390. This nav carries no `.btn` today, so the CSS defect was latent here rather than shipped — the :not(.btn) scoping is in src/shell.css so the next lane to add a header CTA does not inherit an invisible one.
r8: strip HTML comments as their own pass before stripping tags. <[^>]+>
stops at the first >, so a comment containing one is only half-removed. On this artifact a single realistic source comment leaked 93 characters of invisible text into text, which the bare-email scan and the text-floor check both read.
Page-level treatment is ProjectAmp2/agents/SHELL.md. The tokens block in
src/shell.css between TOKENS-START/TOKENS-END is this site's own; everything below TOKENS-END is the shared shell.
amp-nav.js at the repo root is lane N's, fanned out by sync-nav.sh. The
build does not touch it.
This file said "spec + marketing site. No implementation code yet." That was false, and it had been for long enough that a site lane was briefed from it. Measured on 2026-08-16:
mix test | 66 tests, 0 failures (:db tag excluded by default) |
POST https://app.agentelic.com/mcp tools/list | HTTP 200 — 10 tools, not the 7 AGENTS.md calls "planned" |
tools/call template_list | 200 — 2 templates, both hashes reproduced from this tree |
GET https://app.agentelic.com/api/agents | HTTP 500 — the persistence read path is broken on the deployed app |
GET https://app.agentelic.com/ | 404 by design — router.ex declares POST routes only |
| checkout / billing / subscription code | zero matches across lib/, test/, config/ |
So: the MCP surface is live_deployed, the pipeline and test runner are in_tree, staged deployment and anything commercial are spec. The page states those four rungs separately rather than averaging them.
The $49/mo and $199/seat prices are labelled `proposed` and flagged [TRAVIS]. They price accounts and seats that do not exist in the source. Whether Agentelic sells anything is not this page's call — it was not deleted on one, either.