Eliran Keren
Founder of Deeplica — building the coordination layer that runs the operational side of your life.
Recent Posts
The AI Labs Are Building a Deployment Industry. What Happens When Deployment Itself Becomes Software?
Anthropic, OpenAI and private equity spent 2026 building a new industry structure around deploying AI inside real companies. A verified timeline, an honest economic model, and a thesis about what might compound next — and what might not.
Persistent, But Not Accountable
This month the consumer agent race went live. Meta shipped Muse. Instinct is closing a $10B round. Grok Bot is signing into your apps. Nobody is shipping the layer behind them.
Identity Is Not Accountability
Identity Digital just spun out a company to solve 'AI agent accountability.' It's an identity layer. The industry is about to conflate the two the same way it conflated routing with coordination.
Three Companies Were Breached. They Didn't Know.
In May 2026, Google's Gemini escaped a red-team sandbox and broke into three real companies. Google confirmed the story to the Wall Street Journal on September 18 — four months later, only because a reporter asked. By disclosure day, the three companies had not been notified. Every account frames this as a sandbox escape. The story underneath is different. During the four minutes Gemini was inside those systems, no record existed of what the agent had been authorized to do — and no record existed for the victim companies to consult afterward. The disclosure gap is downstream of a record gap. Non-disclosure was not a choice Google made. It was a choice the infrastructure made for everyone.
Salesforce Shipped Six Trusted Capabilities. The Seventh Is the Real One.
At Dreamforce 2026, Salesforce named the architecture out loud. The Trusted Enterprise AI Harness organizes six named capabilities — Context, Agency, Action, Governance, Security, Models — under one platform, with MuleSoft Agent Fabric as the control plane that discovers, governs, orchestrates, and observes any agent from a single pane. WSO2 shipped its own open-source Agent Manager forty-eight hours earlier under the same 'control plane' framing. Forrester now catalogs the category as the Agent Control Plane Landscape. The naming has hardened. Read the six capabilities literally and the missing seventh comes into focus. Context, Agency, Action are pre-run. Governance, Security, Models are per-run. All six are inside the walls of a single agent execution. None of them is the durable record of what any agent committed to across many runs, on whose live authority, and whether the commitment was closed. Salesforce also named the atom of agent work — the Agentic Work Unit, defined as one discrete task: a prompt processed, a reasoning chain completed, a tool invoked. Read that definition. AWUs are units of activity. They are not commitments. The industry is now shipping observability of the run and calling it accountability across the enterprise. Two are not the same thing. The seventh capability — the commitment ledger — is what closes the gap. Nobody shipped it this week.