The hosts of Agentic AI in DevOps make the case that the "second brain" idea — long a personal-productivity meme — is exactly what AI agents need to be useful inside a real engineering environment. Without a system of reference, the agent burns its context window rediscovering what is running where, hallucinates the gaps, and needs credentials it should not have. The episode also marks a pivot: B.O.R.I.S is no longer pitched as a replacement DevOps engineer, but as a context layer for engineering systems — and the acronym now stands for Base of Record for Intelligent Systems. Fernando Gonçalves rejects pure RAG as enough on its own (the relationships are what matter), Andrey Devyatkin turns "do you take notes?" into a hiring question, and the half-hour closes with a token-maxing post-mortem and the reminder that you can outsource the legwork to agents but not the decisions.
Summary
The hosts of Agentic AI in DevOps make the case that the “second brain” idea — long a personal-productivity meme — is exactly what AI agents need to be useful inside a real engineering environment. Without a system of reference, the agent burns its context window rediscovering what is running where, hallucinates the gaps, and needs credentials it should not have. The episode also marks a pivot: B.O.R.I.S is no longer pitched as a replacement DevOps engineer, but as a context layer for engineering systems — and the acronym now stands for Base of Record for Intelligent Systems. Fernando Gonçalves rejects pure RAG as enough on its own (the relationships are what matter), Andrey Devyatkin turns “do you take notes?” into a hiring question, and the half-hour closes with a token-maxing post-mortem and the reminder that you can outsource the legwork to agents but not the decisions.
Key Topics
What a second brain actually is
Fernando opens with the personal version of the concept: as you learn more, holding everything in your head stops working. Information either falls out or fails to connect. A second brain is the external place you put it so you can come back later, find connections you missed the first time, and stop reinventing your own thinking. He admits the discipline used to be hard — you would write notes, never reread them, and forget what they meant. AI changes that, because now the notes can be queried, summarised, and cross-referenced without the manual review step that used to kill the habit.
Vladimir adds a tooling note. The second-brain space used to be a churn of dozens of apps — Apple Notes, Evernote, Obsidian, and many others — with people changing platforms every few months. Vladimir says the trend has moved toward simpler markdown files in the filesystem. Time that used to go into learning the tool now goes into actually writing.
Notes as a hiring signal
Andrey says he used to interview for this directly: Are you taking notes? What’s your system of taking notes and recalling it? The answer told him a lot about how organised the person was and how they wanted to operate. Fernando seconds it — he asked similar questions, and people who wrote things down learned faster and were less likely to redo the same work. The mechanism, both argue, is partly discipline and partly the act of writing itself; even typing solidifies a thought differently than just thinking it.
There is a side joke that with a good enough archive you could ask AI to argue with you in your own voice. Andrey: “I could get a firsthand experience of trying to argue with myself — be in the shoes of my wife for a second.”
The DevOps engineer as help desk
The transition into the work argument is direct. A DevOps engineer is, in practice, a help desk for developers — where is this thing, how does it work, where is it deployed, when was it deployed. AI can answer those questions, but only if it already knows the answers. Otherwise it has to rediscover the environment on every session.
Vladimir makes a related observation: in any organisation, the person being asked is usually the one who has been there longest — the CTO, an early developer, an infrastructure engineer, or whoever has been there longest. It is a natural human pattern, but it does not scale and it does not survive that person leaving.
Why letting the agent rediscover is dangerous
Fernando connects this back to incidents the audience will recognise — the “AI deleted my database” stories that the show has touched on before. If you tell the local agent (Claude Code, Codex, Kilo, take your pick) to go and rediscover the cloud environment, you have to give it credentials. The credentials have to be broad enough to find what you have. That is the danger zone: now an autonomous loop may have write paths it did not really need.
There is also the efficiency cost the show has covered before. Every rediscovery is a round of API calls that eats the context window the agent needs for the actual task, and the longer the agent loops the more likely it gives up and hallucinates. The argument links straight back to the context window discussion in episode #2: tools, skills, and system prompts are the useful part of the context, and a second brain is what stops them from being crowded out by rediscovery.
Why RAG alone is not enough
Andrey is blunt about retrieval-augmented generation, which “people have been running around talking about for four years now.” Vectorise your documents, do similarity search, hand the chunks to the model. It helps; it is better than nothing. It also stops short of what infrastructure work actually needs.
Fernando walks through why with an AWS example. Change a security group rule and the question is not “what is the security group” — it is what traffic flows through it, what is downstream, what is the blast radius. Pure RAG returns the document about the security group. It does not return the dependency graph. A graph-shaped representation does, and the episode lands on graph-RAG as the direction the hosts see as necessary for serious engineering use cases. He admits the team learned this the hard way — they started with pure RAG, built more on top, and only then did the relationship gap become obvious.
Andrey adds two more dimensions the discussion treats as table stakes:
- Temporality of state. What the system looked like today is not what it looked like yesterday. To argue about the system over time, the context store has to represent change, not just current snapshots.
- Staleness vs live truth. Any database goes stale. If the question is about right now, the agent cannot fully trust the stored answer; it needs a hybrid recall that layers live data on top of cached context, revalidating on the fly. Otherwise the answer is confidently wrong.
Both hosts note this is harder than it sounds. “People tell us, isn’t this simple to build? If you want to do it properly, no, it isn’t.”
B.O.R.I.S becomes a context layer
This is the pivot the episode is built around. The earlier framing of B.O.R.I.S — visible on getboris.ai as of the recording on June 5, 2026 — was a DevOps agent: an AI teammate aimed at replacing or augmenting the human DevOps engineer. The team is moving away from that framing.
Andrey’s read is that many teams in 2026 are more interested in building their own agents than buying a ready-made DevOps agent. “Everyone is a builder.” Teams want to assemble their own agents, in their own harnesses, with their own models. So the value B.O.R.I.S delivers is no longer the agent itself; it is the knowledge that any agent needs to actually work in a private engineering environment. Public framework documentation is everywhere — pick a Node.js or Spring stack and any coding agent can write against it. What is missing is the private context: this account, this VPC, these modules, these Slack threads, these decisions. B.O.R.I.S is being repositioned as the layer that supplies that, plugged in over MCP so any harness can consume it.
Vladimir frames the practical benefit the same way. Teams that try to build their own equivalent are starting from scratch — the ingestion, the graph, the temporal layer, the freshness story. Plugging into a context layer that already exists is cheaper than rebuilding it.
He also flags some non-obvious dimensions a real context layer has to carry: not just what is running where, but the team’s conventions, code guidelines, operating principles, and even artefacts that get re-used on a calendar — the example he gives is the annual audit pack. It is many shapes of knowledge living in one graph.
The new acronym
Andrey clarifies the rename. The product name B.O.R.I.S predates the acronym; the hosts joke that it is not named after Boris Grishenko from GoldenEye. As of this episode it stands for Base of Record for Intelligent Systems — the records being the nodes intelligent systems consume. He is candid about how product naming usually goes: you pick the letters, then you grow the meaning. The first acronym fit the DevOps-agent goal; the new one fits the context-layer goal.
Software demand is growing, not shrinking
Fernando opens a tangent the room runs with. The 2024–2025 prediction was that AI coding agents would shrink demand for software engineers. What he is seeing in practice is the opposite: the cost of building software has dropped enough that smaller companies and individuals now want their own custom systems, and that demand may create opportunities outside the headcount squeeze inside big tech. Coding agents are good at the code itself, but the infrastructure to run that code — and to scale it — still needs human judgement. He sketches the consultancy version of the bet: an independent engineer plus a context layer like B.O.R.I.S could support many small environments at a fraction of the manual overhead, instead of starting each one cold.
Token-maxing starts to fade
Andrey reports a related trend from the vendor side. The race to chew through as many tokens as possible — internal leaderboards, “boil the ocean, go go go” — looks to be over. Companies are settling on per-person monthly caps; he names roughly $1500/month/person as a number he has heard. The framing has shifted from hype currency to a budgeted tool.
Vladimir speculates the spike may have been related to IPO/document-filing optics — he flags this as his own guess, not a settled read. He also notes the corollary: when subscriptions get capped, local models become more attractive again. He has been experimenting with them on an older MacBook and expects that to keep getting easier.
The frame both come back to: AI in the engineer’s toolkit is a perk, like a good laptop. Coding with it is still mentally draining work — the cognitive load is real even when the typing isn’t — and the engineer is still in the loop. “You can outsource the legwork, but you cannot outsource the decision making.” Vladimir puts the accountability point bluntly: there always has to be someone responsible. “It can’t be ‘our AI agent decided to tweak the database that way.’”
Andrey closes by pointing listeners at a recent Pragmatic Engineer episode with the co-founder of OpenCode, which he calls a balanced view on what it actually feels like to ship software with AI today.
Resources
- Building a Second Brain — Tiago Forte — the canonical write-up of the personal-productivity concept Fernando opens with: capture, organise, distil, express.
- Obsidian — a markdown-first note tool mentioned during the discussion.
- Model Context Protocol (MCP) — the standard the hosts reference for how a context layer can be plugged into any agent (Claude Code, Codex, Kilo, OpenCode); useful background on how B.O.R.I.S is exposed to consumers.
- Microsoft GraphRAG — open-source implementation of the graph-augmented retrieval pattern the hosts argue is the direction needed for infrastructure context stores; a starting point for understanding why pure RAG misses the relationships that matter.
- B.O.R.I.S — Base of Record for Intelligent Systems — Sirob Technologies’ context layer for engineering systems, repositioned in this episode from a DevOps agent into a knowledge layer that any agent can query.
- Episode #2 — The Tool Layer: What Makes Agentic AI Possible — prior episode on context windows, cross-referenced when explaining why letting the agent rediscover is so expensive.
- Episode #7 — When Agent Memory Helps and When It Hurts — the foundations episode on agent memory; the second-brain framing here sits on top of the procedural / semantic / episodic distinction covered there.
- The Pragmatic Engineer podcast — Andrey’s pointer for a balanced read on building software with AI today, citing a recent episode with Dax Raad, co-founder of OpenCode.