When does an earlier permission stop covering the next action? Dominick connects this week's conversations about agent payments, memory, interfaces and protected acceptance checks.

This episode draws on the October 25 DeltaX Weekly newsletter and four published X Articles. It separates test proposals from results, roadmap items from shipped features, and bounded synthetic demonstrations from general reliability claims.

Narration uses an authorized AI clone of founder Dominick Noval's voice. Grok Imagine visuals are conceptual illustrations, not recordings of actual product operation or real events.

Sources
The Next Action Needs Its Own Evidence: Source
Checkout transition test proposal: Source
Stripe: Treasury for agents: Source
The PR Is Green. Did the Agent's Next Action Change?: Source
DeltaX Preflight public Action preview: Source
DeltaX Preflight MCP v0.1.0: Source
Preflight synthetic walkthrough: Source
Memory write and merge discussion: Source
Are we giving AI agents too much context that they never actually use?: Source
Anthropic: Agent Skills donation to AAIF: Source
The Agent Didn't Get Smarter. Its Interface Did.: Source
Agent S4 paper: Source
Your Agent Found the Site. It Still Couldn't Read It.: Source
Ora: AgentReady launch experiment: Source
OpenAI monitoring incident report: Source
The Agent Failed at Step 40. The Mistake Was Step 12.: Source
Computer-use debugger research: Source
Computer-use debugger project: Source
Agility Robotics: Digit 5 collaboration plans: Source
API Repair Bench evidence and limits: Source
Revised API Repair Bench film: Source
Aishwarya Naresh Reganti: reviewing self-improving harnesses: Source

Chapters
00:00 When permission gets old
00:51 What does revocation stop?
01:56 The name can stay while the effect changes
03:02 Check the write and the merge
04:03 Version the instructions that travel
04:33 Sometimes the interface is the intervention
05:08 Reachable information comes first
06:07 Protect the monitor from the migration
07:10 Find the first wrong move
08:21 The same boundary reaches physical systems
08:54 Keep a fresh run separate from a film
09:30 What DeltaX means here
10:04 Try one changed condition
11:05 What I will watch next

A tool call can succeed while the task remains unfinished. Dominick connects this week's published DeltaX Articles, community discussions and research through three decisions: finding the current brief, knowing what confidence predicts, and checking the work behind a success claim.

October 18, 2026 | DeltaX Weekly with Dominick Noval

Narration uses Dominick Noval's authorized AI voice clone, generated locally. Grok Imagine scenes are generated conceptual illustrations, not actual product operation or real events. Diagrams explain arguments and proposed tests.

Sources
LinkedIn newsletter: Source
More Agents Are Not a Free Speedup: Source
Your Agent Knows the Task Is Hard. Does It Know This Attempt Is Failing?: Source
The Paper Was Finished. The Experiment Wasn't.: Source
The Unread Message Was Part of the Task: Source
PersonalAgentEval: Source
Outcome grounded confidence calibration: Source
Memory fact evolution and FEB-30: Source
Approve What Exactly: Source
Into the Unknown: Source
Vals AI team comparison: Source
GitHub current policy at startup: Source
GitHub agentic autofix: Source
Author calibration explanation: Source
Recovery and an unresolved request: Source
Human correction before older extraction finishes: Source
Separating memory stages: Source
A calibration clarification: Source
The complete approval card: Source
Review unflagged runs too: Source
October 13 API Repair Bench film: Source
Published Repair Bench engineering boundary: Source
Connectome Entity public snapshot: Source

CHAPTERS
00:00 Before we call it done
00:49 The unread requirement
01:43 The order of facts
02:40 What confidence predicts
04:03 Unknown after a timeout
04:35 Approval needs the whole effect
05:37 A report needs the work
06:40 A useful robotics correction
07:07 More workers are an intervention
08:04 Our own public evidence
08:39 What DeltaX means here
09:43 Three questions for a review

When the situation changes, what makes an agent’s next action acceptable? Dominick connects this week’s published DeltaX X Articles, memory research, agent controls and operating-environment evaluations, then looks ahead to SF Tech Week.

October 4, 2026 | DeltaX Weekly with Dominick Noval

Narration uses an authorized AI clone of Dominick Noval’s voice, generated locally. Grok Imagine scenes are AI-generated conceptual illustrations, not real product operation or events. Research results apply to their tested conditions. DeltaX Evaluate remains under restricted verification; public website access is not public API access.

SF Tech Week’s official dates are October 5–11. Dominick’s stated focus is October 5–9. This is forward-looking; no attendance or future conversations are claimed. The October 4 LinkedIn newsletter is published: https://www.linkedin.com/pulse/before-next-action-questions-im-taking-sf-tech-week-dominick-noval-moe6c/

SOURCES

• The Agent's Next Thought Has a Gate — Published DeltaX X Article, September 28
• The Fastest Desktop Can Make the Agent Slower — Published DeltaX X Article, October 1
• The Agent Model Is Not Portable Yet — Published DeltaX X Article, October 2
• Your Agent Guardrail Cannot See the Team — Published DeltaX X Article, October 3
• Open Agent Safety Platform — NVIDIA technical announcement, September 28
• Getting started with Claude Code mods — Claude Developers, October 1
• Authorization — MCP specification
• ORBIT — Research preprint, September 27, discussed this week
• When Context Changes — Research preprint, September 30
• Hiding in Plain Sight — Research preprint, September 30
• cua-speedrun — Research preprint, September 30
• Multi-harness RL — FineEnvs primary results record
• Memory decision models — LinkedIn community discussion, October 3
• Shared reasoning and operational state — LinkedIn community discussion, October 3
• SF Tech Week, October 5–11 — Official event calendar
• Connectome Entity — Public DeltaX research snapshot

October 4 DeltaX Weekly newsletter

What should an agent remember, retrieve, and re-check before acting? Dominick connects this week's DeltaX X Articles and public technical discussions through working memory, cache costs, and the boundary between saved information and current authority.

September 27, 2026 | DeltaX Weekly with Dominick Noval

Narration uses an authorized AI clone of Dominick Noval's voice. Grok Imagine scenes are generated conceptual illustrations, not recordings of product operation or real incidents. Research results belong to their source authors; hypothetical examples and proposed tests are identified as such.

CHAPTERS

00:00 What should an agent carry forward?

00:58 Memory needs a working set

02:41 Events, preferences and constraints

03:10 The cost of the computed past

04:53 Structure and output order

05:28 What survives a boundary

07:06 A promise needs a mechanism

08:11 Evaluate the trajectory

08:48 DeltaX research and changing conditions

10:01 What to test next

SOURCES

DeltaX: Your Agent’s Memory Needs a Working Set

https://x.com/deltaxevaluate/status/2102162154208338371

DeltaX: Parallel AI Doesn’t Need a Clean Slate

https://x.com/deltaxevaluate/status/2102534366853546022

DeltaX: The Agent Loop Has a Cache Tax

https://x.com/deltaxevaluate/status/2103983625687728477

DeltaX newsletter: What Changes When the Agent Can Change Itself?

https://www.linkedin.com/pulse/what-changes-when-agent-can-change-itself-dominick-noval-rmp0c/

SoL-Pi research

https://arxiv.org/abs/2609.20519

Nebius / WEKA vendor benchmark

https://nebius.com/blog/posts/nebius-weka-shared-kv-cache-benchmark-hgx-b300

dQwen3.5 research

https://arxiv.org/abs/2609.20751

OpenAI disclosure: simulated prompt injections

https://alignment.openai.com/misalignment-reports/self-replicating-prompt-injections-exist/

Claude Code plugin evaluations

https://code.claude.com/docs/en/plugin-evals

Reddit: Bots vs Harness

https://www.reddit.com/r/LocalLLaMA/comments/1wr8tf9/bots_vs_harness/

LinkedIn: What Are We Really Evaluating?

https://www.linkedin.com/pulse/what-we-really-evaluating-rethinking-evaluation-ai-agents-mahajan-siduf/

X: plugin evaluation conversation

https://x.com/adocomplete/status/2104019331025883493

DeltaX Connectome Entity public research

https://github.com/DeltaX-Public/deltax-connectome-entity

DeltaX is a dynamic recurrent decision engine. The public Connectome Entity work discussed here is computational simulation research, not a living organism or a physical-world deployment.

YouTube: @DeltaXWeekly

X: @DominickNoval

When agents can produce more work, what makes that work dependable? Dominick connects this week’s DeltaX articles and community conversations through recovery, disputed memory, continuing permission, and the pace of actions that change the outside world.

Episode 1 · September 13, 2026 · Week of September 7–13

CHAPTERS
00:00 When the plan stops matching reality
00:54 Judgment as infrastructure
01:54 Recovery after interruption
03:49 Memory and disagreement
05:40 Permission with context
07:14 Intent as infrastructure
08:21 The pace of external change
09:58 Reports that preserve exceptions
10:57 Research disclosure and evidence
12:04 DeltaX and the week ahead

OUR PUBLISHED ARTICLES
When the Plan Stops Matching Reality — LinkedIn, September 13
Read the LinkedIn article

When Execution Gets Cheap, Judgment Becomes Infrastructure — X, September 7

Intent Is Becoming Infrastructure — X, September 10

AI Safety Is Measuring the Wrong Speed — X, September 12

PRIMARY REFERENCES
MCP authorization specification, version 2025-06-18
Read the MCP specification

EMVCo consultation draft for agentic payments

tldr project — results not independently reproduced for this episode

Anthropic’s September 9 research disclosure

Hypothetical examples and design proposals are identified in the episode. They are not claims that DeltaX has already demonstrated every capability discussed.

PRODUCTION
Narration uses an AI clone of founder Dominick Noval’s voice. Grok Imagine scenes are AI-generated conceptual illustrations, not recordings of product behavior or real incidents. Editorial diagrams are illustrative.

New episodes every Sunday.
YouTube: @DeltaXWeekly
X: @DominickNoval