---
title: "The year our connectors stopped depending on who was online - Atlan Frontier Labs"
url: "https://atlan.com/frontier/essays/connectors-stopped-depending-on-who-was-online/"
description: "Every connector was tied to the specialists who knew it. How Atlan's Connectivity team moved that knowledge into one layer people and agents can both read, and what we learned on the way."
---

An essay from Atlan Frontier Labs (Operating Notes, essay 008) by Mrunmayi Tripathi and Samanvitha Kumar, Connectivity Engineering, Atlan, dated 2026.10.27. How Atlan's Connectivity team moved connector knowledge into one layer people and agents can both read, and what the team learned on the way. Part of [Becoming Frontier](https://atlan.com/frontier/).

## Every Connector Had Its Experts

Until 2026, every connector Atlan ran was tied to the specialists who knew it. A change waited for one of them to be free, and every new connector made the fleet bigger without making anyone freer. Coding agents had not changed that: each engineer got faster at the connectors they knew, but what one of them learnt stayed with them. The bottleneck was where knowledge lived: in a few people's heads, on their laptops, and on dashboards only they watched.

At the start of the year the team committed to working agentically across the whole software lifecycle, as part of [Atlan's wider effort](https://atlan.com/frontier/essays/becoming-a-frontier-company-in-the-open/) to become a frontier company in the open. The first plan, in March, ran the agentic flow end to end on a few connectors first.

Connectors matter beyond the team because of what they feed. Every enterprise AI agent is only as good as the context it reads, and Atlan's bet is a single, neutral [context layer](https://atlan.com/know/what-is-the-enterprise-context-layer/) across every system a business runs. "That layer does not come from a prompt. It comes from a pipeline, and the pipeline is made of connectors." Each connector run is a crawl that extracts a source's metadata (what the data means and where it came from, not just the rows) across warehouses, databases, BI, ETL, streaming and ERP. When a connector fails, the context layer becomes confidently wrong. The team's own shape matched the problem: islands of knowledge and no single layer anyone else, person or agent, could read. In 2026 it built that layer for itself: one picture of the fleet, then rules for every connector, then agents that build connectors.

## Changing the Engine Without Stopping the Car

The first large job handed to agents touched every connector: moving connector workflows from Argo, now the legacy path, to Temporal, which lets the team break a workflow into activities it chooses, each retried and resumed on its own. Agents did the rewriting under one rule: every capability on the old engine had to exist in the rewrite. The rewritten connectors run in production. On comparable runs, Power BI ran nearly five times faster and Tableau more than twice as fast (representative runs: Snowflake 4.7x, Tableau 2.6x, Power BI 4.8x, Databricks 3.1x, BigQuery 4.4x; runtime also depends on the source system and how much changed since the last run). Fresher context followed. With both engines live at once, nobody could see in one place which build was running where.

## One Place to See the Fleet

Connector Pulse, the team's AI-built dashboard, answered that with a view of every deployment across the fleet (live on 6 May 2026), and then kept growing: fleet health across every connector and tenant; sliceable health per customer, connection and failure pattern; views on demand built the same afternoon (an engine-timing view nobody had planned was built with agents in two days); and agent access over [MCP](https://modelcontextprotocol.io/), so agents and people read the fleet from one source.

"Ten failures landing on one customer is one unhappy account. The same ten spread one per customer across ten of them is a fleet-wide pattern. The number is the same, and the response is the opposite."

"Visibility comes before reliability. At fleet scale you cannot improve what you cannot see in one place."

## A Fix Is Spent, a Rule Keeps Paying

With one picture, the same failure became visible twice, and the unit of work moved from the single failure to the class. Every connector is built on the [Connector SDK](https://github.com/atlanhq/application-sdk). A fix is spent the moment it is applied; a rule written into the SDK keeps paying on every connector it touches, every tenant it has not reached, and connectors nobody has built yet. An agent periodically reviews merged fixes and opens pull requests that write fleet-wide patterns into the SDK as conformance rules. An AI reviewer checks each one, then an engineer decides whether it should exist, on purpose, because a wrong rule would push wrong fixes everywhere. Once a rule merges, every connector is checked against it and agents fix most violations on their own.

## From a One-Line Ticket to a Connector

In June 2026 the team filed a ticket reading only "Build SAP Datasphere connector". The agent broke the work into thirteen steps, wrote a specification with 108 traceable requirements, tested against a live instance and found 22 of its own assumptions wrong. It stopped three times for an engineer's call, and by the next afternoon the connector was built, crawled, tested and waiting for review, work that used to take weeks. It is now in [public preview](https://docs.atlan.com/apps/connectors/data-warehouses/sap-datasphere). The agent built on the Connector SDK and shared asset models, years of engineers' knowledge written down where an agent could use it. It settled anything checkable against documentation, the live source or a test, and stopped only for calls a person should make: what the connector is for, which objects matter, and how its assets are modelled. In September, three new Fabric connector capabilities went from request to release in about a week, in time for Microsoft's FabCon Europe.

## We Stopped Waiting for the Customer Ticket

Failure triage used to start with a customer ticket. Now a daily sweep agent reads every fleet failure, often before a customer notices. Agents built on LangGraph work from Pulse's single snapshot of failures:

- One that writes up the day's failures each morning, with one ticket per distinct fault
- One that confirms whether a new release really caused the failures an alert reports
- An agent that takes a support ticket to a root cause and a fix
- One that turns a customer request for a new capability into tested changes
- An agent that watches run times and finds which step is slowing a connector down

The team also watches sources' advance notices of API, deprecation and authentication changes and fixes connectors before they break. A Pattern Analysis Agent view from October 2026 groups each day's failures into patterns, with the fixes the agents opened.

## We Misread Our Own System

Two mistakes cost the most, both about what the agents were given to work from.

- **Two agents agreeing is not proof.** A support ticket carried a lead pointing at the wrong code; the investigating agent and the blind checking agent both started from it, the reviewer agent passed it, and the fix targeted an engine the connector had already left. An engineer caught it from information Pulse shows first. "Two agents reading the same wrong premise had given us one opinion twice." Leads like that are now stripped from tickets, and the support agent checks the engine first.
- **A month's budget in a week.** By the first week of August, agents had spent more than in all of July; about 94% of the spend was re-sending context, with one ticket reaching half a million tokens. The fix: code, not the model, now decides which step of an investigation runs next. Some steps are plain scripts, others get a small or the strongest model, each with its own budget, and a fresh agent sees only what its step needs. Pulse now also tracks each agent's work, outcomes and cost, by graph and by ticket.

## What We're Still Working On

- **Giving every issue an identity.** Each investigation first pins down the connector, connection, run and failure from the system of record, and stops rather than guesses when a field is missing. It works for workflow failures; lineage is harder, because a wrong edge fails nothing.
- **From a right diagnosis to a right fix.** A graph can diagnose correctly and still propose the wrong fix. Misses are closed in the graph itself, and a second agent re-investigates every finished run blind and scores it, so a badly scored run becomes a change to the graph. It is early.
- **Guardrails for a loop that feeds itself.** Agents write fixes and decide which become rules, so a wrong rule could travel as far as a good one. Before agents run more of that loop on their own, the team is building the checks that catch a wrong rule first.

## Where the Knowledge Lives Now

Knowledge and judgement both moved. What an engineer knew now lives in the Connector SDK, in Pulse, and in specialised agents that work across every connector, and a new source can go from an empty ticket to review in about a day. People now decide what a connector should be, and agents do the building.

"Writing a connector is no longer the hard part. Knowing what a source really means, and noticing the day it changes, is what keeps an enterprise's agents from working off yesterday's truth."

The layer the team built for itself is a small version of the one Atlan's customers are building across their whole business. Knowledge compounds once it lives somewhere everyone can read it, and a shared layer earns trust again every day it is used. The next step is a loop that can carry more of the work on its own, once the checks are in place that let us trust it to.

"Every connector we keep true makes our customers' context layer a little more trustworthy for the people and agents who rely on it."

## Author

Mrunmayi Tripathi and Samanvitha Kumar, Connectivity Engineering, Atlan. Atlan's Connectivity team builds and runs the connectors that carry context from each customer's systems into Atlan's context layer. Samanvitha Kumar is Director of Engineering, and Mrunmayi Tripathi leads the team's AI charter.

More essays: [Becoming Frontier](https://atlan.com/frontier/).