Wiring agents to dozens of live source connectors creates fragile chaos; reliable enterprise AI depends on unified permission-filtered indexes, warehouse data modeling, and actively pruned agent memory.
Rolling out an assistant for 10,000 people at Netflix taught Utkarsh Shrivastava that chat interfaces and models change quickly; the durable moat lives below the waterline in governed company context:
“The company context is the key here.”Utkarsh Shrivastava · Platform Challenges Beyond MCP · 00:02:11
When two live connectors answered the same live-event question with different numbers—one from analytics, one from transactional metrics—neither tool was buggy. Adding uncurated MCP servers simply multiplies ways an agent can pick an unauthoritative source:
“You would be very disappointed, but the tool was never wrong.”Utkarsh Shrivastava · Platform Challenges Beyond MCP · 00:04:40
“Is another way for the agent to choose wrong.”Utkarsh Shrivastava · Platform Challenges Beyond MCP · 00:04:40
“The M C Ps have made it very cheap to make a lot of information accessible, but.”Utkarsh Shrivastava · Platform Challenges Beyond MCP · 00:06:49
Evidence: Netflix transcript 00:02:11, 00:04:40, 00:06:49
George Fraser points out that data teams already spent a decade solving inconsistent definitions for human analysts. Wiring an AI assistant directly to raw operational APIs fails immediately on multi-system questions:
“And you can't even do joins across sources, right?”George Fraser · The new job of the data team · 00:07:04
The high-leverage move is pointing agents at the modeled warehouse tables—and adding a compact AGENTS metadata schema acting as a database map of tables, columns, and safe joins:
“They are using as a source of truth for AI agents for questions about data.”George Fraser · The new job of the data team · 00:03:18
“You're presenting a simplified view of what is probably some very messy and complicated data.”George Fraser · The new job of the data team · 00:05:23
“Local tables in a special schema, a map of the rest of the database that acts as a guide.”George Fraser · The new job of the data team · 00:11:18
Evidence: Fivetran transcript 00:03:18, 00:05:23, 00:07:04, 00:11:18
For unstructured knowledge across Slack, Jira, Drive, and GitHub, Aamir Shakir shows why one-MCP-per-tool breaks down: the agent wastes turns guessing where documents live, and hits the slowest or most rate-limited API in the chain:
“Like getting the right data and access to data is actually the bottleneck, and not intelligence.”Aamir Shakir · How to build agents over your own data · 00:03:12
“And then your agent has the task to guess, basically, what might be the right tool containing this information.”Aamir Shakir · How to build agents over your own data · 00:02:08
“We always say you are only as good as your data, but in this case, you're only as good as your weakest connector.”Aamir Shakir · How to build agents over your own data · 00:04:07
Preparing one unified search layer upfront—with document ACLs attached at ingest—transformed a banking assistant’s accuracy from ~40% to 97%, echoed by Parallel’s requirement to return source citations on every grounded lookup:
“The answer accuracy jumped from very forty percent to ninety-seven percent”Aamir Shakir · How to build agents over your own data · 00:14:19
“Not only do we give you the answer, we also give you the citation, so you can go fact check the source yourself.”Devin Gupta · The first Grounded Decision Model · 00:05:45
Evidence: Mixedbread transcript 00:02:08, 00:03:12, 00:04:07, 00:14:19
Context engineering governs software repositories as tightly as data warehouses. Lauren Tan warns that coding agents treat the existing repo as their primary prompt—replicating messy hacks unless teams keep patterns clean:
“It's now likely to be just copied by the next agent, since it looks like the established pattern in the codebase.”Lauren Tan · All Code Is Slop (Except Mine) · 00:04:26
Three builders model the surrounding decision layer explicitly: Ankit Jain captures why a change happened directly from agent sessions; Thariq Shihipar maintains a linted, English-like spec layer above the agent-written engine; and Fabian Hedin stores task execution in an append-only, Git-like trajectory:
“Intent is why are we making whatever change.”Ankit Jain · One team's path to trusting AI generated code · 00:08:38
“It's supposed to feel like English. It's kind of like a spec, and I'll show you what it looks like. But it's at the density where, if I just read this code, I know what's happening, and I can edit it. I can tweak it. But more importantly, I just like you know can have a bit of a ground truth of like what the what my game looks like.”Thariq Shihipar · Thariq Shihipar, Independent · 00:13:27
“So unlike it, we cannot merge agent trajectories”Fabian Hedin · What it takes to set intelligence free · 00:06:56
Unattended agents that append notes forever degrade into noisy prompts. Vertical AI builder Mihail Eric promotes customer context via structured reflections, while Overcut scores retrospective memories so low-value traces never re-enter context:
“This is where we developed a concept called reflections”Mihail Eric · Learnings from Monaco · 00:10:23
“We are scoring them, so the next run we run only on top of the memories that have the highest score.”Raz Berman Gold · The Software Factory That Needs Zero FDEs · 00:07:30
Lance Martin formalizes this at frontier scale: stronger models know when to read and write memory natively, yet greedy in-session notes still conflict globally—requiring an offline dreaming pass that compiles transcripts into structured wikis (outperforming semantic search alone on the Stainless corpus):
“This is a fundamental paradox of agent memory because agents write memory typically as they go about a task.”Lance Martin · Exploring Claude’s frontier · 00:07:06
“They know intrinsically when to write to memory, and they kind of know intrinsically when to read from memory to retrieve information.”Lance Martin · Exploring Claude’s frontier · 00:08:54
“Dreaming lets you reflect over transcripts and these written notes, and re-represent them.”Lance Martin · Exploring Claude’s frontier · 00:12:13
“The wiki gets better and better and better, allowing Claude answers questions.”Lance Martin · Exploring Claude’s frontier · 00:15:15
Evidence: Anthropic transcript 00:07:06, 00:08:54, 00:12:13, 00:15:15
While teams agree modeled context beats raw connector sprawl, two counter-claims define the operational limits of passive logs and unified read indexes:
“So you just gotta go in there and just aggressively slash the crap out of it.”Steve Yegge · Escape from Trash Mountain · 00:10:42
“So to be able to do that correctly, you need to give right amount of access to all of these agents.”Karan Vaidya · Run a Company on Agent Swarms · 00:05:49
Arsalan Bashir (Sentry) bridges read modeling and live API execution: stop handing agents 1:1 raw REST endpoints where they confuse similar tools and loop on pagination. Wrap cross-source ID resolution and deterministic joins inside clean use-case tools:
“The agent starts picking the wrong tools sometimes.”Arsalan Bashir · Let the Agent Cook · 00:01:32
“You know, organize tools by a clear use case.”Arsalan Bashir · Let the Agent Cook · 00:03:01
00:03:18, 00:14:19, 00:05:49).00:03:01).00:07:30, 00:12:13, 00:10:42).Connecting company data to a shared AI assistant is now cheap; the platform's real job is deciding which source answers, who owns that answer, and how hard to gate by stakes, with evals in CI and company context as the lasting investment.
A vertical AI company is a shared harness (data, an agent workspace with memory, expert-guided workflows and evaluation) plus domain nuance, and it wins by putting its best domain experts inside the product and tracing everything, because correctness is whatever those experts say it is.
Once AI writes most of the code, a team should stop reviewing diffs and instead capture intent, decisions and evidence, owning the rules that verify the code rather than the code itself.
The data stack built for business intelligence is now the context layer for AI, so data teams should connect agents to modeled warehouse data rather than wiring them to every source system.
Agent-written code can improve in quality at hundreds of PRs a day only if every correction is encoded into the codebase as a deterministic check and review shifts to verification artifacts; automating a sloppy loop just produces more slop.
Agents over company data should search one index prepared ahead of time, with access rules attached, not query every source live through one connector per tool, because the bottleneck is context, not intelligence.
A year of software factories shows fuel is the hard limit and currently subsidized, agents behave like staff who need managing and pruning, and the more agents you run the more humans you need, so spend the fuel on experiences.
Overcut argues that an enterprise software factory should not need forward-deployed engineers: it plugs into the tools and agents you already run, composes specialized governed agents into SDLC workflows with human gates, and improves itself through scored retrospective memories.
Because Lovable's output is a business, not code, letting an agent work for ten hours needed a runtime that survives deploys, computers that start and fork in milliseconds, and two verification loops, the outer one fed by business metrics Lovable owns.
Claude's frontier is self-organisation: a thousand Fable 5.2 agents given only git and a SQL database built their own coordination to formalise Fermat's Last Theorem, and stronger models write, read and, through dreaming, consolidate memory into better representations.
To build agents that investigate data on their own, consolidate tools by use case and push deterministic complexity into them, stop pre-chewing context the model can now handle itself, and always give the agent a cheap way to check its own work.
Rather than acting as an alien intelligence that forces developers into a single narrow interface like Bash or React, improving models form a broad 'smooth interior' of capabilities that enables a software Renaissance—provided builders pair agentic engines with human-readable specs to stay in creative control.
Grounding decision models with live web search from a large proprietary index solves uncalibrated overconfidence in ungrounded LLMs, delivering higher accuracy, verifiable citations, sub-second latency, and one-dollar-per-thousand-request pricing.
Running agent swarms across a company means giving each one scoped access to every app it needs and a firewall on what goes in and what it does, because attackers already use agents.