Chain of Thought | AI Agents, Infrastructure & Engineering

Cisco is giving each of its 90,000 employees a personal AI agent. Some of those same employees are already connecting agents from outside the security boundary without asking IT. DJ Sampath, SVP and GM of AI Software and Platform at Cisco, walks through what had to be in place before the rollout: scanning every prompt, sandboxing each agent, keeping agents multi-tenant, scanning the skills and tools people bring in, and settling which identities an agent may act as.
New episodes, the ideas behind them and Conor's essays land in the Chain of Thought newsletter first. Subscribe: https://newsletter.chainofthought.show/
We recorded at Cisco's San Francisco office, a block or two from OpenAI, where Cisco was an early Codex design partner. DJ explains how engineers rework their skills files with each model generation, how per-engineer PR bots review agent-written code, and why defenders now have to move at agent speed when an exploit can show up within minutes of a CVE going public.
We cover:
What Cisco's IT team required before rolling out MyAgent to 90,000+ employees
Prompt injection by email, and why the first integration everyone makes is the riskiest
Cisco's AI security taxonomy, DefenseClaw, and Project Quarry, a new open source language for describing agent behavior and policy
How the agents are actually used: daily briefs, meeting actions, live dashboards, 2 trillion tokens in a single week
Why skills files need refactoring every model generation
Discovering and governing shadow agents with zero trust access
Multi-model code review, spec drift, and evals built from pull request traces
How Cisco Cloud Control was built in about eight months with agentic coding
Chapters:
(0:00) Guardrails tested inside Fortune 100 companies
(0:21) Securing 90,000 employee AI agents
(8:04) Cisco's open source agent security stack
(12:01) Project Quarry: a policy language for agents
(17:07) Personal agent use cases at Cisco
(19:08) Early design partner on Codex
(23:15) Reassessing workflows each model generation
(27:27) Shadow agents and zero trust access
(30:29) Security as an adoption accelerator
(31:39) Multi-model code review for dependability
(36:09) Defenders moving at agent speed
(41:04) How Cisco Cloud Control was built
Connect with DJ Sampath:
X: https://x.com/djsampath
LinkedIn: https://www.linkedin.com/in/djsampath/
DJ on Cisco Blogs: https://blogs.cisco.com/author/djsam
Project Quarry: https://blogs.cisco.com/ai/project-quarry-agent-security-detections
Cisco: https://www.cisco.com/
Related episode: Cisco President and CPO Jeetu Patel, "Parenting Your AI Agents for Best Results": https://share.transistor.fm/s/43e97aa6
Connect with Chain of Thought host Conor Bronsdon:
Newsletter: https://newsletter.chainofthought.show/
Twitter/X: https://x.com/ConorBronsdon
LinkedIn: https://www.linkedin.com/in/conorbronsdon/
YouTube: https://www.youtube.com/@ConorBronsdon
More episodes: https://chainofthought.show

Thanks to Inngest, presenting sponsor of season four of Chain of Thought. Agents in production run long: they call models and wait on APIs and people. But the longer agents run, the more they break. Inngest handles that with durable execution. You build your agent as steps in TypeScript, Python, or Go. When a step fails, Inngest retries it with exponential backoff, and completed steps are saved and skipped. Try it out: https://inngest.link/cot-pod
Thanks to Walrus, presenting sponsor of season four of Chain of Thought. Walrus Memory gives AI agents portable, verifiable memory that carries context across apps, sessions, and other agents. Get started at https://walrus.xyz/cot
Thanks to WisdomAI for sponsoring this episode. In WisdomAI's latest benchmark, Claude Code alone topped out at 73% accuracy; WisdomAI closed the gap, delivering accurate answers 3x faster at about a third of the cost. Get "The Blueprint for Trusted AI Analytics": https://wisdom.ai/benchmark
Thanks to Svix, presenting sponsor of season four of Chain of Thought. Svix delivers billions of reliable webhooks for startups and the Fortune 500. Get started at https://link.svix.com/cot. Qualified startups get $12,000 in credits, and YC companies get $50,000.

Andy Triedman is a partner at Theory Ventures, and he is building his own software factory, while assessing how startups are leveraging theirs. His read after running agent fleets: the models are frankly good enough for most of the work. The vast majority of things you'd ask an agent to do are not intelligence limited. They are context limited. We have infrastructure work to do - but we can be more ambitious!
Full transcript and chapters: https://chainofthought.show/podcast/79-for-ai-agent-fleets-the-model-is-rarely-the-bottleneck-theory-ventures-andy-triedman/
So the hard part moves upstream. Deciding what the agents should build, what they need to know, and when they should stop and ask a person is now most of the job.
We cover:
Moving from solving one problem at a time to building systems that solve whole classes of problems, like generating and testing integrations
Which tasks still need human approval, and which can be fully handed off
Why a short feature spec leaves too much unsaid, and how better questions up front cut review interruptions
Keeping shared context and company decisions current, because an agent takes everything it reads as true
Inference ROI, and when well-scoped tasks can go to cheaper models
Sandboxes and simulated customer environments for agents
First steps for teams: shared skills, cloud agents and approval gates
Protecting customer data, and organizing agent memory from session traces to company priorities
New episodes, the ideas behind them and Conor's essays land in the Chain of Thought newsletter first. Subscribe: https://newsletter.chainofthought.show/
Chapters:
(0:00) Agents are context limited, not intelligence limited
(0:34) From coding agents to software factories
(5:53) What still needs human approval
(10:16) Agent loops that improve the product on their own
(14:52) Where agent automation breaks down
(17:59) The bitter lesson and engineering as experimentation
(20:08) From writing code to directing agent fleets
(24:18) Shared context and stale decisions
(27:25) Inference ROI and cheaper models
(30:23) Moving human time from code to customer intent
(34:42) Sandboxes and simulated customer environments
(37:26) First steps: shared skills and cloud agents
(40:14) Protecting customer data and isolating agents
(42:32) Context limits and hierarchical agent memory
(47:05) What Theory Ventures looks for in founders
Connect with Andy Triedman:
LinkedIn: https://www.linkedin.com/in/andy-triedman/
Andy at Theory Ventures: https://theoryvc.com/team-members/andy-triedman
Theory Ventures: https://theoryvc.com/
Connect with Chain of Thought host Conor Bronsdon:
Newsletter: https://newsletter.chainofthought.show/
Twitter/X: https://x.com/ConorBronsdon
LinkedIn: https://www.linkedin.com/in/conorbronsdon/
YouTube: https://www.youtube.com/@ConorBronsdon
More episodes: https://chainofthought.show

Thanks to Svix, presenting sponsor of season four of Chain of Thought. Svix delivers billions of reliable webhooks for startups and the Fortune 500. Get started at https://link.svix.com/cot. Qualified startups get $12,000 in credits, and YC companies get $50,000.
Thanks to Walrus, presenting sponsor of season four of Chain of Thought. Walrus Memory gives AI agents portable, verifiable memory that carries context across apps, sessions, and other agents. Get started at https://walrus.xyz/cot

Thanks to Inngest, presenting sponsor of season four of Chain of Thought. Agents in production run long - they call models and wait on APIs and people. But the longer agents run, the more they break. Inngest handles that with durable execution. You build your agent as steps in TypeScript, Python, or Go. When a step fails, Inngest retries it with exponential backoff, and completed steps are saved and skipped. Try it out: https://inngest.link/cot-pod
Thanks to G2i for sponsoring this episode - for over a decade, they vetted and placed engineers at other companies, from startups to FAANG. Two years ago, they turned that same judgment inward, building their own bench to review RL environments, evals, and training data that models are trained on. Get access: https://fandf.co/3SFxVm6

What breaks when agents become the customer? Most of the integration layer, APIs, SDKs and webhooks alike, was built for human developers and for servers that are always listening. Agents are neither.
Full transcript and chapters: https://chainofthought.show/podcast/78-what-breaks-when-agents-become-the-customer-svix-ceo-tom-hacohen/
When an agent runs on a laptop, closing the lid takes down the tunnel its webhooks come through. When the lid opens, it gets a storm of retries from 12 hours earlier. Svix founder and CEO Tom Hacohen's answer is to let the agent poll for what it missed.
Svix delivers webhooks for companies like Clerk and Resend, and it created Standard Webhooks, the spec OpenAI, Anthropic and Google Gemini have adopted. In this episode of Chain of Thought, Tom joins Conor Bronsdon to talk through designing APIs for agents that adapt, whether agents still need SDKs, and why he wants less code even when agents write it. He also introduces Poke, Svix's new spec for making event-driven APIs work for agents.
This episode is sponsored by Svix.
We cover:
Why events are harder for agents than API calls: the sender controls them, and they arrive outside the agent's session
How polling endpoints and Svix's plugins for Hermes and OpenClaw let an agent pick up where it left off
Why Tom calls JSON an abomination and webhooks duct tape, and why worse is better anyway
Why Svix created Standard Webhooks when everyone implemented webhooks differently, and how the spec let Kong and ngrok add webhook verification
Whether agents still need SDKs, and why Svix's skills work like a solutions engineer while its docs stay written for humans
Why Tom wants less code even when agents write it, and what piling patches on patches costs
Poke, Svix's early spec for letting agents subscribe to and trigger events across services, released so feedback can shape it while it's still malleable
Chapters:
(0:00) Duct tape, retry storms and the case for simplicity
(1:19) Why integrations got easier and harder at once
(2:59) How EteSync's webhook requests led to Svix
(5:22) Worse is better: why webhooks persist
(7:15) What changes when an agent receives the event
(9:07) What happens when the laptop closes
(11:08) Hermes, OpenClaw and agents that poll
(13:03) Letting agents choose their own subscriptions
(14:42) The long tail of the web without APIs
(18:38) Why Svix created Standard Webhooks
(24:17) How Standard Webhooks is governed
(26:16) Following real usage, including MCP
(28:17) Designing APIs for agents that adapt
(30:38) Do agents still need SDKs?
(31:25) Skills as documentation for agents
(34:52) Less code, even when agents write it
(36:19) Agentic development and the offshoring analogy
(37:34) Software factories and the plumbing underneath
(40:57) Keeping up while the standards change
(44:18) What Svix is launching with Poke
(46:39) Why Svix released the Poke spec early
(47:58) Building for your customers' customers
(49:38) Clerk, Resend and webhooks in the background
(52:22) Where to follow Tom
Links from the episode:
Svix: https://link.svix.com/cot
Poke: https://www.svix.com/poke/
Standard Webhooks: https://www.standardwebhooks.com
Standard Webhooks on GitHub: https://github.com/standard-webhooks/standard-webhooks
EteSync: https://www.etesync.com
ssot-check: https://github.com/conorbronsdon/ssot-check
Most of the Web Will Never Get APIs for AI Agents | Dhruv Batra: https://chainofthought.show/podcast/63-most-of-the-web-will-never-get-apis-for-ai-agents-dhruv-batra/
Connect with Tom Hacohen:
LinkedIn: https://www.linkedin.com/in/tomhacohen/
X: https://x.com/TomHacohen
Svix: https://link.svix.com/cot
Connect with Chain of Thought host Conor Bronsdon:
Newsletter: https://newsletter.chainofthought.show/
Twitter/X: https://x.com/ConorBronsdon
LinkedIn: https://www.linkedin.com/in/conorbronsdon/
YouTube: https://www.youtube.com/@ConorBronsdon
More episodes: https://chainofthought.show
Thanks to Svix, presenting sponsor of season four of Chain of Thought. Svix delivers billions of reliable webhooks for startups and the Fortune 500. Get started at https://link.svix.com/cot. Qualified startups will get $12,000 in credits, $50,000 for YC companies.

If your agent's memory lives inside one model provider, switching models or harnesses means starting over. You need portable memory - yet when switching agent memory between models, you can swing accuracy by thirteen points or more, depending on which model reads them back.
Full transcript and chapters: https://chainofthought.show/podcast/77-switching-models-shouldnt-mean-starting-over-walrus-protocols-kimberly-logan/
Kimberly Logan, Head of Product at the Walrus Foundation, where she builds Walrus Memory, joins Chain of Thought to discuss how to build portable and verifiable memory for AI agents - and why Walrus's decentralized storage network has proved an apt primitive. She and host Conor Bronsdon get into why model lock-in is a memory problem, what people use agent memory for beyond coding, how you prove what happened to your data, and why Walrus doesn't publish memory quality benchmarks, plus more. This episode is sponsored by Walrus.
We cover:
Why longer context windows don't help once you switch models or harnesses, and why an import function only fixes one point in time
Why Kimberly doesn't benchmark Walrus Memory against memory startups, and why she sees the opportunity as bigger than memory alone
What people build with Walrus Memory besides coding agents, which Walrus puts at maybe 30% of usage: long-running study aids, trading agents, and security teams comparing production snapshots with code changes
Walrus's figure that almost 85% of the memories written are read back more than 30 days later
How attestations on the blockchain and a delegate key you control let you show what happened to your data, and why unchanged data still isn't proof that it's accurate
Why Kimberly treats recall quality and latency as table stakes, and how a knowledge graph keeps recall on the freshest memories
Why she left almost 20 years in traditional software, including seven at Google, for a blockchain company, and where shared memory across companies and trust boundaries goes next
Walrus Console, announced this week: one place to manage agent memories and files through an interface or an MCP connection, without knowing how the storage underneath is configured
Connect with Kimberly Logan:
LinkedIn: https://www.linkedin.com/in/krlgn/
Twitter/X: https://x.com/kimblgn
Walrus: https://walrus.xyz/cot
Connect with Chain of Thought host Conor Bronsdon:
Newsletter: https://newsletter.chainofthought.show/
Twitter/X: https://x.com/ConorBronsdon
LinkedIn: https://www.linkedin.com/in/conorbronsdon/
YouTube: https://www.youtube.com/@ConorBronsdon
🔗 More episodes: https://chainofthought.show
Chapters:
(0:00) Why agent memory has to move with you
(1:25) Context windows don't follow you to a new model
(4:41) Why an import function doesn't fix lock-in
(6:38) Why Walrus isn't competing head-on with memory startups
(8:39) Where Walrus Memory goes next
(10:38) Trading agents and security snapshots on Walrus Memory
(14:44) Verifiability: answering what happened to your data
(17:21) Fresh recall, and why Walrus doesn't publish memory benchmarks
(20:22) From traditional software to a blockchain company
(24:40) A record other companies and regulators can check
(25:42) What Walrus Console is for
(29:09) Use agent memory beyond coding agents
(30:52) Plugging in your own agent stack
(31:41) Memory across companies and trust boundaries
(35:48) Closing thoughts
Thanks to Walrus, presenting sponsor of season four of Chain of Thought. Walrus Memory gives AI agents portable, verifiable memory that carries context across apps, sessions, and other agents. Get started at https://walrus.xyz/cot
#AI #AIAgents #AgentMemory #SoftwareEngineering #DevTools

Andrea Malagodi is CTO at Sonar, which builds AI code verification and governance tools. When finance asked about the size of his team's AI bills, he calculated cost per PR. At one extreme was an engineer with more than 500 PRs in a short period and a low cost per PR. At the other was an engineer with very high spend and very few PRs.
Full transcript and chapters: https://chainofthought.show/podcast/76-your-best-ai-engineer-might-have-the-worst-metrics-sonar-cto-andrea-malagodi/
Judged on cost alone, the first engineer wins and the second looks like a problem. Malagodi looked closer. The first had built a personal agent factory, with specification, design review, documentation records, coding, verification and functional testing, and about two-thirds of the output was tests and validations. The second was working on a hard problem that needed the AI to reason through long, multi-turn sessions and didn't reduce to a line count.
Numbers alone tell you something, but not which engineer was doing the more valuable work.
We cover:
Why cost per PR is becoming the industry's standard AI spend metric, and how Malagodi's 500-PR engineer and high-spend engineer broke it
The history of bad proxies, from lines of code to story points to token leaderboards, and why Malagodi says the people who removed code were often the most important
Sonar's guide, verify, solve loop: Vortex for codebase context, algorithmic plus LLM-based PR analysis, a remediation agent for legacy debt, and the Hunter agent for security vulnerabilities
Why Malagodi breaks agent work into small pieces, and why a five-day session that produces 400,000 lines of code leaves you with no idea what is in your codebase
Setting up disagreement between agents with an orchestrator and personas for engineering, product management and quality, plus Conor Bronsdon's cross-model-family review setup
Rolling AI tools out at enterprise scale: one cost view for finance, up to 30% savings from cutting repeated context reads, and why budgets as small as $50 a week are hard to work with
Rotating a triage duty to protect deep work, including how Sonar's teams built skills to sort through hundreds of reported CVEs, a number of them likely false positives
New episodes, the ideas behind them and Conor's essays land in the Chain of Thought newsletter first. Subscribe: https://newsletter.chainofthought.show/
Chapters:
(0:00) Human reviewers rubber-stamp big AI changes
(2:43) Learning the craft alongside AI tools
(9:09) From lines of code to PR counts
(12:17) Spreading top engineers' setups org-wide
(16:50) The guide, verify, solve agent loop
(22:51) Specialized agent lanes and multitasking limits
(26:51) Clear asks and guardrails for agent coordination
(30:29) Building disagreement between agents
(34:59) Mixing models instead of picking one
(39:46) Rolling out AI tooling at enterprise scale
(46:19) Defense in depth for AI-written code
(50:52) Protecting time for incident investigations
(54:08) Personalized software for teams and individuals
Links from the episode:
Sonar Vortex research: https://www.sonarsource.com/blog/cut-your-coding-agents-cost-with-sonar-semantic-code-navigation/
SonarQube Hunter Agent: https://www.sonarsource.com/products/sonarqube/hunter-agent/
Connect with Andrea Malagodi:
LinkedIn: https://www.linkedin.com/in/malagodia/
Sonar: https://www.sonar.com/
Connect with Chain of Thought host Conor Bronsdon:
Newsletter: https://newsletter.chainofthought.show/
Twitter/X: https://x.com/ConorBronsdon
LinkedIn: https://www.linkedin.com/in/conorbronsdon/
YouTube: https://www.youtube.com/@ConorBronsdon
More episodes: https://chainofthought.show

Thanks to Svix, presenting sponsor of season four of Chain of Thought. Svix delivers billions of reliable webhooks for startups and the Fortune 500. Get started at https://link.svix.com/cot. Qualified startups get $12,000 in credits, and YC companies get $50,000.
Thanks to Walrus, presenting sponsor of season four of Chain of Thought. Walrus Memory gives AI agents portable, verifiable memory that carries context across apps, sessions, and other agents. Get started at https://walrus.xyz/cot

Thanks to Inngest, presenting sponsor of season four of Chain of Thought. Agents in production run long - they call models and wait on APIs and people. But the longer agents run, the more they break. Inngest handles that with durable execution. You build your agent as steps in TypeScript, Python, or Go. When a step fails, Inngest retries it with exponential backoff, and completed steps are saved and skipped. Try it out: https://inngest.link/cot-pod
Thanks to G2i for sponsoring this episode - for over a decade, they vetted and placed engineers at other companies, from startups to FAANG. Two years ago, they turned that same judgment inward, building their own bench to review RL environments, evals, and training data that models are trained on. Get access: https://fandf.co/3SFxVm6