Agentic Engineering Live: rUv in London · Thu 15 Oct 2026 · 18:00 to 21:00 BST · Agentics Foundation London chapter · hybrid
The talk is scripted: slides, builds, the terminal and the live app play on a clock, and the speaker speaks to them. Every number on screen is read from this site's data or from a recorded run. Ten minutes, with a five-minute cut if the slot shrinks.
The player needs no build and no network once loaded. The offline files carry the timeline, the data and the images inside one HTML file; the live-app beat shows a recorded screenshot there, because a page opened from a file cannot load WASM.
The player, embedded: the same file as the full-screen version, playing in this page so the script, the sources and the slides sit side by side. Click into it, then Space to play, arrows or a swipe to move, review to see every slide finished; ⤢ opens it in its own window at the same second.
One card per beat: swipe sideways with a finger (or use the arrows), and scroll the page as usual. Read from the same timeline the player plays.
From the organisers' welcome deck for Agentic Engineering Live (Agentics Foundation, London chapter), Thursday 15 October 2026, Westminster Room, Whitehall House, 41 Whitehall, London SW1A 2BX. Livestreamed and recorded for YouTube: nothing on screen that should not be public.
| time | what |
|---|---|
| 5:30 pm | Arrivals |
| 5:45 pm | Welcome and about the evening |
| 6:00 pm | Agentic Engineering, keynote by Reuven Cohen (rUv) |
| 6:40 pm | Questions for rUv |
| 7:00 pm | Autonomy with a mandate and a memory, Dinis Cruz · 10 minutes, the first of three community presentations |
| 7:10 pm | Community presentation 2 |
| 7:20 pm | Community presentation 3 |
| 7:30 pm | Networking |
How the deck introduces the talk: "Making rUv's agent stack trustworthy: a written mandate for every agent, plus private, verifiable memory, shown across six working proofs of concept."
Builders first, but far from only builders: software and IT services (13 organisations), AI and agentic start-ups (8), consulting and advisory (7), the Agentics community (4), education and research (3), hardware and manufacturing (3), plus finance, cybersecurity, telecoms, engineering, VC, accessibility and design. Counts as the deck gives them, from the work email addresses on Luma.
What that means for the ten minutes. rUv will have spent forty minutes on his stack just before: skip re-introducing it and point back to his keynote. Half the room does not write code: lead each beat with what changes (a written mandate, evidence you can check) before how. The consultants, the bank and the security people will ask about audit: the receipts and the run vault are the answer.
Open the player on the laptop. Press A: an audience window opens. Drag it to the projector and press F there. The laptop now shows the presenter view; both windows share one clock.
Press Space (or the clicker's play key) when you start speaking. The beats change on time without a click.
Space pauses the clock. → or the clicker's next (PageDown) jumps to the next beat; ← goes to the start of this beat, twice quickly for the previous one. B blacks out the projector.
The player types can the coder send email? into the live page itself. If you click into the page to type something else, press Esc to give the keyboard back to the player.
offline-10.html: same talk, no network.rehearsal.json; drop it below to see the drift.npx caches the package).All keys: Space play/pause · →/PageDown next · ←/PageUp back · P presenter · A audience window · F fullscreen · B black · V review (no animation) · T theme · R rehearsal · 0 restart · ? help · click the progress bar to seek.
Rendered from the timeline file the player plays, so the script and the slides cannot drift. Print this page for a paper copy of the words.
Rule 2 of this build: verify before you claim. The table and ABP slides are not typed into the timeline; the player reads them from the site's data at load time. The terminal beats show commands and the output their recorded runs produced.
| On screen | Value | Source |
|---|---|---|
Ruflo's default for email.send | allowed, legacy-default-allow | evidence/ruflo-policy-legacy.json (ruflo 3.56.10, fresh project) |
| The coder's ABP rows | 10 rows, owner, valid until | data/abp.json: 6 roles, 36 rows, 0 gaps |
| Same 11 actions, default vs ABP | 6 stopped (4 denied, 2 approval), 5 allowed | data/report.json (POC1) |
| Ledger | valid, 11 receipts each side | data/report.json · data/state-abp.json · evidence/poc1-run.txt |
| Memory history and restore | 4 commits; coder and tester entries back after a fresh clone | data/poc2-run.txt (local roundtrip; the CI jobs are wired and need the secrets) |
| RVM rights per role | 6 roles, GRANT never delegated, 0 gaps | data/rvm-caps.md (POC3) |
| RuVector's own boundaries | witnesses "do not encrypt memory content"; filters "are not a complete authorization boundary"; a delete "does not automatically remove copies in snapshots, branches, replicas, exports" | RuVector README, "Security and governance" (github.com/ruvnet/ruvector, main branch, read 15 Oct 2026) (quoted, not run) |
| The live app | RuVector WASM 2.3.1, 36 rows indexed | app/, also in the public vault okbiml09 |
The terminal shows the command and the output the real run produced, captured beforehand and kept in the repository. Anyone can run it again from Replicate and compare. The live app beat is live.
No. Its CLI wants an interactive terminal for policy admin; POC1 uses the library engine, its documented programmatic surface. The Claude Code session (POC6) is a runbook for a human at the terminal.
You can. The ABP is the layer above: one mandate per deployment with the why and the barrier per row, compiled to Ruflo rules, RVM rights and CI gates. Change the mandate, recompile all three.
No. sgit encrypts on the client (AES-256-GCM); the server stores ciphertext. The public vault publishes a read key on purpose; the memory vault is private.
It ran locally end to end (two writers, a fresh clone, same answer). The CI jobs are wired on separate runners and need the vault secrets; say so plainly.
A mandate a person can read and sign off per agent, enforced by the engine, and a hash-chained receipt for every decision, kept in a vault the host cannot read. That is evidence for an audit, not a certification.
No. These are proofs of concept using pinned npm packages (MIT, by rUv / ruvnet). The ask is to pick one and build it together.
Press R in the player at the start of a run-through and again at the end: it saves rehearsal.json with the planned and actual start of every beat. Drop it here to see the drift. A beat that starts more than ten seconds late is flagged; the same check runs from the command line with python3 site/build/rehearsal_check.py rehearsal.json.
A voice memo by Dinis Cruz on 14 Oct 2026, written up as The scripted talk on the sgit newsroom: a talk as a timeline that plays on a clock. It grows out of two talks published as vaults: AI vs. AI at Black Hat Europe 2025 (slides as data) and the VoiceDebrief pitch (per-slide timings and a presenter). The first player was a prototype by the newsroom's agent; this one adds the audience window kept in sync with the presenter, clicker keys, a black screen, table rows read from the site's data, a live page that types its own question, an offline build with the images inside, and a timeline check that fails loudly if the beats do not add up.