---
title: "The GitHub for Context Doesn't Exist Yet"
url: "https://atlan.com/context-and-chaos/issue/the-github-for-context-doesnt-exist-yet/"
description: "Two eras of building agents inside my own company, five stack migrations, and the infrastructure question left standing: where is the GitHub for context?"
keywords: "Context Engineering, AI Agents, Enterprise AI, AI Strategy"
---

> Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/

A Context & Chaos deep dive by Prukalpa Sankar (Co-founder & Co-CEO, Atlan), published July 16, 2026 (15 min read), based on her AI Engineer World's Fair talk. Atlan could build an agent in about five minutes; giving it enough business context to be accurate took forever. Across two eras of building agents internally and five agent-stack configurations in twelve months (at one point roughly 300 skills powering 40 active agents), the lesson was that the agents were never the hard part. **The agent layer is becoming disposable. Your company's context is not**, yet most tools still bind the two together.

Backdrop: models are getting exponentially smarter, not exponentially more useful. About one in five organizations reports significant value from AI (Gartner), and [56% of CEOs told PwC](https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-global-ceo-survey.html) they saw neither higher revenue nor lower costs from AI in the past year. Performance is intelligence times context, multiplicatively. Full diagnosis: [the Knowledge Graph Conference talk](https://atlan.com/context-and-chaos/issue/the-context-layer-knowledge-graphs-second-act/) and [the moat piece](https://atlan.com/context-and-chaos/issue/if-intelligence-is-abundant-what-is-the-moat/).

## What the agent needs to know

Maya is a data analyst at McContext Burgers. Marcus, a franchisee with fourteen stores, asks: "Why is my drive-thru time up this week?" Answering needs:

- **Knowledge** (the map of the business): whose definition of "drive-thru time" applies, finance's or operations'? Does "this week" mean Monday to Sunday or trailing seven days, in which time zone?
- **Expertise** (the unwritten diagnostic sequence): Q3 spikes are seasonal and a product launched last quarter, so rule those out first.
- **Norms**: who is asking, whether Marcus needs a metric, explanation or decision, who may see the data, whose definition governs when finance and operations disagree.

Maya learned this by shadowing, mistakes, feedback and edge cases. The engineering question: how do you build the agent Maya?

## Era one: an agent for every job

Eighteen months ago Atlan mapped one team's jobs to be done, judged which AI could do (documentation and meeting prep: yes; relationship management: not soon) and built an agent per job. It hit four walls:

1. **Accuracy.** Agent quality tracked context quality almost exactly; every gap taught a stakeholder not to trust it.
2. **Dependency management and silos.** When marketing changed positioning, the website agent kept pitching the old positioning for weeks. Agents had none of the alignment mechanisms human teams have, and Atlan couldn't see how shipped agents related.
3. **Root cause analysis and traceability.** A bad answer couldn't be traced to model, agent design or context.
4. **Sprawl and drift.** Every agent had its own memory and learned separately; two agents could give three answers.

Underneath: stack churn. Five configurations in twelve months, from a no-code builder to hyperscaler and enterprise-search frameworks to general-purpose coding agents to Atlan's own system of claws in Slack. Customer success and sales used different frameworks, so customer context didn't move between them and agents duplicated work. The framework was disposable; the context was what kept getting lost.

## Era two: one brain, many agents

General-purpose agents like Claude Code inverted the architecture: domain experts encoded knowledge into shared skills in one company brain, and retrieval and assembly connected it to whichever agent was working. Modeled on strong human teams: shared language, a picture of what is true today, playbooks, decision norms, memory of what failed.

- Marketing ran the first experiment: business systems on one side, deliberately heterogeneous agents (coding agents in the terminal, a chat-channel agent, external products) on the other, shared context between.
- The best SEO person wrote the SEO skill; the best competitive-intelligence person wrote that skill.
- An autonomous ads agent needed to know which tables to query (a data graph), then semantics (what ARR means here; what counts as a qualified lead when marketing and finance differ).
- Over six months: about 300 skills and 40 agents (an active point-in-time snapshot). A skill is an encoded procedure another person can use in a general-purpose agent; an agent proactively performs work without waiting for human intervention.

## The wall at 300 skills

- Marketing changed the core category narrative, which fed sales and customer-success email workflows. The narrative had been hardcoded as an earlier snapshot, nothing declared the dependency, and downstream agents kept sending the old story.
- Context must be managed like code, versioned, tested and owned ([argued earlier](https://atlan.com/context-and-chaos/issue/how-to-make-satya-nadellas-vision-a-reality/)). Skills drifted without clear owners or known blast radius.
- Security and governance became "a nightmare": secrets hardcoded in environment files while people downloaded public skill repositories into company workflows. Atlan stopped the affected work and ran a full audit; no breach is claimed. The context layer had become executable infrastructure before it had matching controls.

Era two ran on git from day one, and git is a big part of why it worked. **But git versions text, not meaning.** It can show a definition changed, but not that a new qualified-lead definition changes a battle card two dependencies downstream, that a competitive update contradicts another team's positioning and needs their approval, or turn what 40 agents learned this week into reviewable proposals. Software needed version control plus review, ownership, CI, dependency management, security scanning, package registries and runtime observability. **Context has the diffs. It does not yet have a coherent version of the rest.**

## What the missing infrastructure must do

"The GitHub for context" is shorthand for an operating layer (not necessarily one product) that treats context as durable infrastructure rather than agent-specific configuration. Five requirements:

1. **A GitHub-like profile for every unit of context:** owner, approvers, maintainers, scope, provenance, declared dependencies. Skills, semantic models and tools will all be managed as units of context.
2. **Local and company-wide context stay distinct,** with a governed path to preserve legitimate differences (marketing vs finance definitions) and promote local knowledge to shared standards.
3. **Semantic review of change:** which decisions, evaluations and downstream agents a change may affect.
4. **A controlled learning loop:** Atlan experimented with a harness that reads production traces, reverse-constructs candidate improvements and returns them to maintainers as approve-or-reject proposals, with provenance checks, evaluations, bounded promotion and human approval.
5. **Native quality and security:** secrets scanning, permission boundaries, evaluation, rollback, retention and deletion before an agent executes context.

All portable across agent interfaces (MCP, SQL, vector retrieval, hybrid assembly) to avoid recreating lock-in. Full architecture: [the June field guide](https://atlan.com/context-and-chaos/issue/what-an-enterprise-context-layer-actually-is/). Small teams with a handful of agents may only need explicit files, owners and evaluations; the machinery becomes necessary when context crosses teams, changes independently, feeds autonomous work or must survive an agent-layer change.

## Context is IP

If content was king at the dawn of the internet, context is king in the agentic era; the sharper formulation is **context is IP**. Building on Satya Nadella's point that [proprietary knowledge](https://x.com/satyanadella/status/2076323181154230284) is a moat, the defensible asset is the company's accumulated operating model: definitions, exceptions, playbooks, judgment calls and norms, plus the provenance and permissions that make them usable. Competitors can call the same models; context is what they can't download. Context trapped in an agent framework couples operating knowledge to a disposable layer. PK's claim from the stage: this problem is under-hyped, because the hype is on the intelligence axis while the failure is on the other one.

## The five-minute audit

**List every place each production agent gets its context.** For each source, name the owner, scope, downstream consumers, last change, and how you would detect a bad change. A coherent map is the start of a context layer; a pile of agent-specific memories, prompts, repositories and connectors is islands, and the reconciliation bill is accruing.

Signals from the fair: context engineering got [its own track](https://www.ai.engineer/worldsfair/schedule?track=Context+Engineering), alongside memory and graphs tracks, at a conference where [organizers stopped counting around six thousand attendees](https://dev.to/hanzla/inside-ai-engineer-worlds-fair-2026-what-6000-engineers-showed-up-to-build-fgm); one main-stage speaker held the title "Context Engineer"; [one recap](https://getunblocked.com/blog/context-grew-up-at-the-worlds-fair/) and [Latent Space's trends piece](https://www.latent.space/p/aiewf26trends) named the systems around agents, skills included, as defining the year.

Watch the full twenty-minute talk, "WTF Is the Context Layer?": https://atlan.com/know/the-missing-infrastructure-for-production-agents/

## About Context & Chaos

A community newsletter where practitioners, builders and thinkers share lessons on context engineering, governance, architecture, discovery and the human side of data and AI work. Contribute: https://atlan.com/context-and-chaos/contribute/

Related reads:

- [The Context Layer: Knowledge Graph's second act](https://atlan.com/context-and-chaos/issue/the-context-layer-knowledge-graphs-second-act/)
- [What an Enterprise Context Layer Actually Is](https://atlan.com/context-and-chaos/issue/what-an-enterprise-context-layer-actually-is/)
- [If Intelligence Is Abundant, What Is the Moat?](https://atlan.com/context-and-chaos/issue/if-intelligence-is-abundant-what-is-the-moat/)
- [The Onboarding Gap is Killing your AI Agents](https://atlan.com/context-and-chaos/issue/the-onboarding-gap-is-killing-your-ai-agents/)
- [Browse all Context & Chaos articles](https://atlan.com/context-and-chaos/)