Articles about
agentic ai

Claude Enterprise now compares an organisation's settings with a recommended configuration, ChatGPT Enterprise turns dots on by default after an admin preview, and Claude Desktop changes what imports and plugins carry with them. Administration is catching up with agents, and the useful controls are baselines, dated defaults and what moves between accounts.

The first monthly issue of Patterns That Matter to Me reads October's MarTech and AI Watch editions against a register of eighteen patterns. The agent's identity absorbed MCP as the question vendors answer first, skills became a shared format with a check before they run, and Salesforce agreed to buy Braze.

Why I built MarTech Ledger, a dated and source-linked record of what 15 CEP, CDP and B2B marketing automation platforms release, announce and retire, how its AI pipeline works, which tools run it and who it is for.

Claude Desktop can now scan a skill before anyone runs it, ChatGPT Enterprise is moving reusable workflows from custom GPTs to plugins, and Google's agent platform applies the same governance policies whichever model an agent uses. The controls that matter are moving outside the model, to the points where a procedure, a tool call or a connection enters the system.

Anthropic handed Agent Skills to the same open foundation that holds the Model Context Protocol, Google gave skills a place in the Gemini app, and OpenAI's models are on their way into Google's agent tools. A procedure written as a skill is no longer tied to one assistant, which makes the procedure, and who reviews it, the thing worth governing.

Salesforce has agreed to acquire Braze, with the close expected in Salesforce's next fiscal year. For teams running or choosing either platform, the useful question is which parts of their own architecture stay theirs whoever owns the platform: the contracts at the edges, the identifiers that travel, and the instructions written for agents.

Claude Desktop removed old configuration key names and documented that a profile still using them now behaves as if the setting were absent. Google gave Gemini Enterprise admins a date before Gemini 4 Argon switches on and paused skills in Workspace, and Claude's unified view arrives for every organisation unless a key is set. Configuration for agentic tools now has a shelf life.

GPT-6 with Intelligent UI began rolling out to every ChatGPT tier, so an answer can arrive as a chart, a form or a small tool. Around it, OpenAI and Anthropic published two different scopes for marking AI text, Google Docs started editing Markdown natively, and Claude Desktop put a border around imported chat history.

Insider made agent autonomy a mode chosen per use case and channel, 6sense let workflows be rehearsed in a sandbox before promotion, Iterable spelled out the order in which a control must be switched on, and Marketo's MCP server gained smart list operations during its open beta. Rollout is becoming part of the product.

How Codex, Claude Cowork and I shared the video factory work, using written handoffs, independent checks and human decisions to turn each correction into something the next production could reuse.

OpenAI's dots arrive with read-only background research and allow, ask or block rules, and its specialist dots with an identity of their own, while Anthropic, Google and Braze tightened what an agent may do under a person's name. Who the agent is decides what the audit log can tell you.

Braze put its Operator inside Claude and ChatGPT under the individual user's identity, Marketo's MCP server runs as a dedicated API user, and OpenAI gave its specialist dots an identity of their own. Three answers to the first architecture question of agentic marketing.

Claude Desktop, Google Workspace and Gemini Enterprise put an expiry date or a new owner on standing access, while two ChatGPT defaults moved towards fewer prompts. Why agent governance is shifting from a one-time approval to a recurring one.

Iterable made the email tracking pixel a per-recipient consent value and Braze made its MCP server sign in as the person using it. The same design decision from two ends of the stack, and the history nobody has planned to keep.

One vendor removed a confirmation prompt by default and shipped the largest set of enterprise controls yet in the same release. Another published six reports on things its models did wrong. Both belong in the same conversation.

Salesforce and HubSpot both announced that their own screen is no longer where you use them, and both answered the governance question with inheritance. What that leaves unresolved for anyone who has to audit it.

How repeated editorial work on this site became named procedures with their own caps, order of operations and approval gates, what that changed about consistency, and where the comparison to journey orchestration honestly stops holding.

In one release train, three separate misconfigurations stopped being ignored and started stopping the product. Fail-closed arrived in the admin console, and a frontier model shipped switched off.

Three customer engagement platforms published what their agent may actually do, and they put the answer in three different places: the launch copy, a reference table, and the API contract. The ones that write the boundary down are the ones you can buy from.

Six real failures from my own site, none of which were wrong answers. Truncated links that passed every build, a translation that was grammatically perfect and about the wrong thing, and what they say about where governance actually lives in an agentic operation.

Optimizely gave its agents their own identity. Salesforce and Anthropic gave theirs yours. Salesforce is also turning the agent platform on for you by default. Treasure AI published a read-only boundary, Salesforce retracted a write capability it had documented, and the FTC sued Amazon over an auction that was not the auction it described.

Last edition closed on a vendor describing its organisation-wide instruction setting as guidance the model follows, not an enforced control. Six days later, Anthropic and Google both shipped the enforcement column.

Every agentic demo shows one task in one session. This is what an agentic operation actually consists of on month five, using my own site as the evidence: a written context layer, a memory with expiry, and one unambiguous source of truth.

In two weeks the agent acquired a trigger it does not control, a credentialled session it can reuse, and a schedule nobody attends. Then OpenAI pointed the same architecture at the workspace itself and, unlike everyone else, wrote down what it may not do.

In eight days Anthropic shipped scanning for third-party skills, Google shipped an administrator approval queue for them, and Treasure AI redefined a saved workflow as a skill. Software supply chain vocabulary, applied to prose. Meanwhile the read-only phase of agent integration ended without an announcement.

Three agent platforms published releases in the same week and shipped no new model, no new agent type and no new autonomous behaviour. What they shipped instead was enterprise identity, tool-call tracing and consumption billing, which are the three things that were actually blocking rollouts.

One frontier model landed this week, and it was not the story. The consequential releases were an administrator key that makes MCP tool approvals expire with the session, and a console for finding out which agents you are already running.

Adobe took the motivation layer of loyalty into the CEP and left the points ledger behind. Two vendors shipped remote MCP servers and sold the boundary rather than the reach. A week where every new surface arrived with its constraint attached.

The first edition of a new weekly series on agentic work platforms. This week the big three stopped selling capability and started selling control: OpenAI Presence, cryptographic agent identity in Gemini Enterprise, and governed publishing in Microsoft Copilot.

In a single week, 6sense put its intelligence behind MCP, Iterable pulled live data into journeys, and Adobe Journey Optimizer shipped both. Three vendors, one move: the line between knowing the customer and acting on them is giving way.

Why LLM agents grounded by feedback, memory and constraints feel so unsettlingly familiar, and what that means for MarTech architecture.

Marketo brings MCP to the MAP layer, Iterable ships the unglamorous fundamentals, Mailchimp complicates its own strategic signal

MCP is moving from isolated platform news to a common access layer for customer engagement platforms.

Agentic AI promises to make MarTech platforms easier to use, but without the right architecture and operating model it moves complexity rather than removing it.