Adam Jensen

I've seen a lot of announcements in the last few weeks about "Agent Mesh" or "Agent Fabric", most recently Mulesoft's announcement yesterday. It's got me thinking. What these products attempt to do - at least on paper - is to package together many of the "outbound" controls of AI Gateways that govern how software can interact with LLMs, like prompt guards, PII sanitization, and cost management, with the "inbound" controls like authentication, MCP server management, and orchestration through A2A, into a single opinionated toolchain. And it Is this a good idea? Are we ready for it yet as an industry? I'm not sure. On the one hand, any software that leverages LLMs will always have special requirements - it's non-deterministic, and usually expensive. And is often being granted considerable levels of privilege. Some controls are obviously necessary. And it's certainly not hard to imagine a world of discrete, re-usable agentic services that will require platform-wide solutions for authentication, observability and composition. And... I get it, all these vendors are under a lot of pressure to demonstrate that their "platform" will solve for AI use-cases both now and in the future. It reminded me a lot of the Service Mesh craze that kicked off in 2017. And I think there are some important lessons we can learn - namely that it's very hard to get governance right in a single platform that works for everyone. The major service mesh projects like Istio, Linkerd or Kuma were built because their engineers had felt a decade of real-world pain orchestrating micro-services. Even so, after the hype wore off, the category prompted a major backlash as developers and platform teams charged with actually solving problems felt that these Mesh products did more to get in their way and slow them down than they did actually solve any problems. 8 years later, one could argue that only now is Service Mesh emerging as a category useful enough to effectively solve real problems in real-world deployments. Agent Mesh products, by contrast, haven't been forged from real-world pain. And very few organizations are truly at a point yet where agent orchestration and portfolio management is a real problem they need to solve. I'm interested to see where this category goes, but my feeling is that it will take a while to get good enough to be really useful, and there will be a lot of bumps along the way. In the meantime if you are building AI, make sure you get the basic stuff right first. AI Gateways and API Gateways (and yes, Service Mesh) are solving real problems today - and can set you up very well for an agentic future. Don't expect a Mesh or a Fabric to solve your AI problems for you.

37

2 Comments

Every PMM team needs to do this now! Get access to the source code repo. Here is why: 1. Ask questions in Claude or Cursor 2. Uncover how the product works 3. Answer technical and high level questions quick 4. Format outputs by audience 5. Even draw flow diagrams Now there is never a new question you can’t answer in minutes. No need to ask a PM. No need to wait.

12

In a new blog post, Marc Bowes looks at how updates work in Aurora DSQL, and what it means for scalable schema design. Read it here: https://lnkd.in/gmGEK8Dv A couple take-aways for application builders: - You mostly don't have to worry about hot read keys in DSQL, even in read-write transactions. DSQL can scale out per-key read throughput, and the design requires no co-ordination between transactions reading the same key. (Thank you, MVCC and Time Sync!) - Scalable applications do need to avoid high-contention write keys when doing updates. Writes need to be coordinated and ordered (to implement the I in ACID), which limits per-key throughput. - Most applications can avoid hot write keys with the right schema design. Most applications have already been written this way, because high-contention keys perform poorly on relational databases of all kinds.

240

11 Comments

💥 I just found a foundational flaw in Claude. All models, all interfaces. Repro steps below: 1. Go to Claude settings 2. Attempt to update your email address 3. If you can’t find how to do it on your own, ask Claude how to change your Claude email address 4. Validate that Claude is not hallucinating by going to https://support.claude.com and searching for “How do I change the email address associated with my account?” 5. Shake your head and think “sorry users, I know we kind of slid past this when you signed up. Hope the email you used to sign up never gets compromised or needs to change – magic links or SSO are the only option (darn it). And if it does, hope you didn’t buy an annual PRO plan because no refunds / transfers even if to the same human ✌ (grrrr…)” Perhaps this behavior is the result of conscious prioritization based on a thorough understanding of needs / trade-offs. If this behavior was an oversight, then based on prior trips on the modernization train, this could be a massive slog of tech debt. Particularly in a hyper-growth business dependent on uninterrupted & high-integrity data access / entitlements / provenance. Product people – let's use this moment to build foresight! - Don’t abdicate PM or UX from “technical” discussions about data models, identity, or access control – decisions within these domains have incredible implications - Customer onboarding & admin workflows are foundational CX, not afterthoughts - If you’re using Claude Code, invest in Git + GitHub integration for improved resilience (and much more)

3

We were stuck debating roadmap priorities. Instead of guessing, I interviewed 16 customers. Turns out, that “small” usability fix we’d deprioritized was actually blocking enterprise renewals. We built it. Renewal rate jumped 14%. Lesson: The loudest signal doesn’t always come from leadership. It usually comes from your customer. #VoiceOfCustomer #CustomerInsight #B2BSaaS

32

6 Comments

What separates people who get great results from Claude from those who don't? It's not what you think. I've spent hundreds of hours building production agents on Claude, Copilot Studio, Claude Code, raw API calls. Same model, wildly different results depending on how you interact with it. Three patterns changed everything for me. 1. Edit, don't reply When Claude gives you broken code, don't reply with "that didn't work." That keeps the failed attempt in context. Claude anchors on it and tries to patch its own mistake instead of rethinking. Edit your original prompt instead. Add a constraint like: Fix the root cause in [function_name] without adding wrapper functions. This removes the bad attempt entirely. Clean context, sharper constraint, better output. Fewer tokens burned. 2. Ban workarounds up front Claude's default instinct when code fails is to make it stop failing. Not the same as making it work correctly. It adds try/catch wrappers. Fallback values. Defensive layers that hide the real bug. In agent workflows, this is lethal. A silently failing tool call cascades through the entire chain. Put this in your system prompt: Do not add fallbacks to mask errors. If core logic fails, surface the error. Fix the underlying logic only. This one instruction changes Claude's behavior more than anything else I've tried. 3. Plan before you build In agentic coding tools, Claude can enter rapid edit-run-edit loops that compound mistakes. Each iteration adds context and moves further from intent. Require a plan first. Review it before approving changes. If the plan says "add a fallback," reject it. If it touches unrelated files, push back. Making Claude's reasoning legible before it acts is the single highest-leverage habit you can build. Bonus Pin your constraints at the top of the prompt. Instructions at the beginning and end of a long context window get more weight than those buried in the middle. "No fallback logic" on line 3 hits harder than the same rule on line 200. None of this is about clever prompt tricks. It's context engineering. You're controlling what's in the context window, constraining the model's solution space, and making its reasoning legible before it acts. The model isn't the bottleneck. The interaction pattern is. #LegalTech #AI #ContextEngineering #LLM #AgenticAI

11

1 Comment

Having run out of things to say, said no one ever, Bradley Shimmin and I roll out his latest brainchild, "the callout" segment on our Agents of Dev podcast. Brad recognizes Microsoft for its GA release of Semantic Link, which allows you (and now your agents) to build and work with semantic models in Power BI. Brad does an illustrious job of describing the origins of semantic models in BI and why they are so vital to agents and their development. Amen, brother. It's a great lead into other topics on this episode, including decoding what is happening at Entire, and semantic models underpin so much of my work defining our Agent Control Plane framework. So much more in this episode. Please check it out, follow, and share the podcast with your friends. Link is in the comments. The Futurum Group DevOps.com Techstrong.ai

18

2 Comments

Stop sprawling DX. The 2025 report confirms what operators feel. Leaders fund outcomes, not slogans. Three goals: productivity up, legacy out, CX up. Big blockers: silos, surprise costs, missing skills. So, we shipped one route in Sonoma, not twelve. Stack: Cursor + Lovable + n8n + Manus AI + Supabase with Twilio. Route: webhook → validate → enrich → schedule. Act: send SMS, create job, assign owner. Why this fits the report: reduce debt, center data, measure. Agents that can’t write back aren’t agents. Upskilling matters: we ran a six-week AI + automation sprint. Silos fall when every update lands in one table. Costs drop when glue code stays tiny. Governance: retries, limits, audit logs by default. Goal stays simple: 95% pass before we scale. Pick one process → Measure → Optimize. No fluff, just ROI.

32

10 Comments