> ../signals/2026-11-13.md
The last few weeks were less about new models and more about the layer around agents: who watches them, who lets them in, and who owns the door to your systems. Five items worth your time, and what each means if you are building or buying agents.
§ 01 · CrowdStrike buys Fiddler for agent observability
On November 3 CrowdStrike announced a definitive agreement to acquire Fiddler AI and fold it into Falcon as a product called Falcon Observe, with general availability planned for its fiscal Q4 2027. The pitch is tracing what an agent did, which tools it called and why, before and after deployment. CrowdStrike announced Dreadnode, an AI security testing company, the same day. Source: CrowdStrike.
What it means: "what did the agent do and why" is becoming a security question, not only an engineering one. If you cannot answer it today from your own logs, start there before you buy anything. Our notes on instrumentation cover the minimum.
── ◆ ──
§ 02 · Five9 agrees to buy Level AI for about $325 million
On November 5 Five9 signed a definitive agreement to acquire Level AI, a customer experience AI company, for approximately $325 million in cash. Level AI's applications will be sold inside Five9's platform and as standalone products for contact centers on other systems. Source: Five9 8-K.
What it means: contact center AI is consolidating into the big platforms. For a mid-size team that is good news for features and bad news for flexibility. The question of build or buy for support agents gets more important, not less.
── ◆ ──
§ 03 · Cloudflare adds analytics and access policies for personal agents
On November 4 Cloudflare launched a Personal Assistants category in Security Analytics and per-agent access settings (Allow, Block on pages with ads, Block) for supported personal agents. The setting is now called AI Crawler and Agent access policies. Source: Cloudflare.
What it means: if your agent browses the web on someone's behalf, more sites will now decide explicitly whether to let it in. Build for being blocked. We wrote about this in the verified agent pattern and agents that get empty pages back.
── ◆ ──
§ 04 · Uber's MCP Gateway: 800+ servers, 5,000+ tools, one door
Uber described the MCP Gateway it built to standardize how its agents discover and call internal services. It hosts over 800 MCP servers and over 5,000 tools, and supports incremental discovery so agents do not load thousands of tool definitions into context. Sources: Uber Engineering, InfoQ.
What it means: you do not need Uber's scale to use the idea. Even with five agents, one controlled entry point for tools beats each agent wiring its own. We wrote it up as the tool gateway pattern.
── ◆ ──
§ 05 · Google puts a general agent inside Gemini Enterprise
On October 8 Google announced a Gemini agent for enterprise users that takes on multi-step work tasks from the Gemini Enterprise app. Source: TechCrunch.
What it means: general office agents are now bundled with the suites companies already pay for. Custom agents earn their place on the processes those general agents do not reach: your order desk, your invoices, your rules.
── End of signal ──
One pattern across all five: the value is moving from what the agent can do to how it is watched, limited and let in. That is where most production failures we see come from too.
ORBIRESEARCH