Research: Context as Infrastructure in GenAI Collaboration

Date: 2026-03-10 Search queries used:

• “context as infrastructure GenAI collaboration knowledge management”
• “context engineering LLM persistent context sessions AI workflow”
• “context window management AI collaboration professional knowledge workers designers”
• “"context as infrastructure" OR "contextual infrastructure" AI language model”
• “extended mind theory Andy Clark distributed cognition context AI collaboration philosophy”
• “knowledge work context continuity cross-session AI collaboration design strategy professionals”
• “context amnesia LLM stateless sessions professional knowledge workers continuity rituals artefacts”

Executive Summary

Context as infrastructure is the practice of treating the information that grounds an AI collaboration — sources, histories, constraints, domain knowledge, role definitions — as a maintained, versioned system rather than an ad-hoc input regenerated per session. The term has moved from a design metaphor into an operational reality: Anthropic (2025) articulated context engineering as the discipline replacing prompt engineering; enterprise data platforms (Atlan 2026) now sell dedicated context layer infrastructure; and practitioners are developing rituals and artefacts to bridge the gap between sessions. The core problem is “context amnesia” — the stateless nature of LLM sessions forces professionals to rebuild shared understanding from scratch, creating a “translation tax” that erodes the value of AI collaboration over time. For knowledge workers in design, strategy, and product roles, the question is not just technical (how to persist context across API calls) but also epistemic and organisational: what context is worth maintaining, who owns it, and how does it evolve without drifting from current reality.

Key Sources

Effective Context Engineering for AI Agents — Anthropic

• URL: https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
• Type: Technical blog / practitioner guide (Anthropic Applied AI team, Sep 2025)
• Key points:
  • Defines context engineering as “the set of strategies for curating and maintaining the optimal set of tokens during LLM inference”
  • Identifies “context rot”: performance degrades as context window fills, even with large windows
  • Recommends compaction, structured note-taking, and sub-agent architectures for long-horizon tasks
  • Advocates “just-in-time” context retrieval: lightweight identifiers (file paths, links) loaded on demand, mirroring human cognition
  • Distinguishes operational context (high-velocity task state) from decision context (stable codified knowledge)
• Tenet alignment: Directly embodies the Context as Infrastructure tenet; supports Human Intent First (context keeps intent legible across sessions)
• Quote: “Treating context as a precious, finite resource will remain central to building reliable, effective agents.”

Beyond Prompts: Why Context Engineering Is the Next Big Shift in AI — Centizen Nationwide / Medium

• URL: https://medium.com/@centizennationwide/beyond-prompts-why-context-engineering-is-the-next-big-shift-in-ai-eee131688e4c
• Type: Industry analysis article (Jan 2026)
• Key points:
  • Frames operationalisation as: “treat context as infrastructure, not an ad-hoc prompt file”
  • Identifies the bottleneck: “the bottleneck isn’t the model size, but how well they assemble and refresh context”
  • Governance angle: context pipelines must provide audit trails, privacy filters, data lineage tracking
  • Predicts context engineering will become foundational enterprise infrastructure within 12–18 months
• Tenet alignment: Aligns with Context as Infrastructure; governance angle touches Pluralism (context must include multiple frames, not just dominant organisational view)
• Quote: “Prompt engineering set the intent; context engineering builds the environment where that intent thrives.”

Mastering the Context Workflow — Ja’dan Johnson / Medium

• URL: https://jadanjohnson.medium.com/mastering-the-context-workflow-a-practical-guide-to-human-ai-collaboration-fa481722dba3
• Type: Practitioner reflection (Jun–Jul 2025)
• Key points:
  • Coins “translation tax”: the cognitive effort of converting human intent into AI-compatible instructions
  • Identifies “context amnesia”: every session starts from scratch, forcing rebuilding of shared understanding
  • Proposes context workflow as: accumulation, role consistency, persistent collaboration across projects
  • Goal is “maximum context transfer with minimum cognitive overhead”
  • Distinguishes automation layer (AI handles independently) from manual layer (humans control directly)
• Tenet alignment: Strongly aligns with Human Intent First (lowering the translation tax preserves intent clarity); Symbiotic Intelligence (division of automation vs. manual layers)
• Quote: “A great context workflow flips this dynamic. The core premise is designing systems where context accumulates, roles remain consistent, and the AI becomes a persistent & effortless collaborator.”

How to Implement an Enterprise Context Layer for AI — Atlan

• URL: https://atlan.com/know/how-to-implement-enterprise-context-layer-for-ai/
• Type: Enterprise technology guide (Feb 2026)
• Key points:
  • Describes four storage patterns: graph databases, vector stores, rules engines, temporal databases
  • Separates layers: ingestion/extraction → semantic translation → governance/policy → retrieval/activation
  • Emphasises continuous feedback loops: every AI interaction refines the context for future interactions
  • Context drift: “models fail when training data conflicts with current reality” — temporal mismatch
  • Workday quote: “All of the work that we did to get to a shared language amongst people at Workday can be leveraged by AI via Atlan’s MCP server.”
  • Mastercard: “We have moved from privacy by design to data by design to now context by design.”
• Tenet alignment: Strongly aligned with Context as Infrastructure; governance layer aligns with Pluralism; feedback loops align with Always Scalable
• Quote: “Organizations succeeding with production AI treat context as infrastructure rather than application-specific configuration.”

The Extended Mind — Andy Clark & David Chalmers

• URL: https://www.alice.id.tue.nl/references/clark-chalmers-1998.pdf (original 1998 paper)
• Type: Academic philosophy (Analysis, 1998; foundational)
• Key points:
  • Proposes “active externalism”: cognitive processes can extend beyond the brain into tools and environment
  • Argues that if an external tool reliably plays the same functional role as an internal cognitive process, it is part of the cognitive system
  • Otto’s notebook example: an external notebook functions as long-term memory when reliably consulted
  • Coupling matters: the tool must be actually used, not merely available
• Tenet alignment: Philosophical foundation for Context as Infrastructure — context systems ARE cognitive infrastructure; supports Symbiotic Intelligence (genuine extension, not mere automation)
• Quote: “The human organism is linked with an external entity in a two-way interaction, creating a coupled system that can be seen as a cognitive system in its own right.”

Context Engineering: Sessions, Memory — Patrick Mcclung / Medium

• URL: https://medium.com/@patmcc1979/context-engineering-sessions-memory-0505b825a273
• Type: Technical overview (2025)
• Key points:
  • Distinguishes short-term memory (in-session) from long-term memory (persisted to external store)
  • Session objects in agent SDKs maintain conversation history; external memory tools retrieve across sessions
  • Memory systems for AI agents are analogous to human episodic vs. semantic memory
• Tenet alignment: Neutral technically, but aligns with Context as Infrastructure when the systems are designed for professional continuity rather than mere task completion

Escaping Context Amnesia: Practical Strategies for Long-Running AI Agents — Hadi Javeed

• URL: https://hadijaveed.me/2025/11/26/escaping-context-amnesia-ai-agents/
• Type: Technical practitioner blog (Nov 2025)
• Key points:
  • Names “context amnesia” as the core failure mode for agentic systems
  • Strategies: memory banks, structured checkpoints, milestone summaries, context refresh protocols
  • Emphasises that effective memory design requires deciding what to remember — selective retention is a craft
• Tenet alignment: Aligns with Context as Infrastructure; selective retention question connects to Human Intent First (remember what serves the human’s intent, not everything)

Major Positions

Position 1: Context as Technical Infrastructure (Engineering View)

• Proponents: Anthropic Applied AI team, Atlan, enterprise AI teams
• Core claim: Context is a system resource that must be architected, versioned, and governed — analogous to a database or API layer. It consists of storage systems (knowledge graphs, vector DBs, semantic layers) plus delivery mechanisms (RAG, context window optimisation, progressive disclosure).
• Key arguments:
  • Context rot means more tokens ≠ better performance — curation is required
  • Just-in-time retrieval mirrors human cognition: we don’t memorise everything, we index and retrieve on demand
  • Enterprise AI that lacks persistent context produces hallucinations and compliance failures
  • Compaction (summarising long context) and structured note-taking (external memory) extend the effective horizon
• Relation to site tenets: Directly instantiates Context as Infrastructure tenet. The engineering view risks reducing context to technical tokens, losing the human-intent dimension unless explicitly governed.

Position 2: Context as Cognitive Prosthesis (Extended Mind View)

• Proponents: Andy Clark, David Chalmers (1998); Annie Murphy Paul (The Extended Mind, 2021); constructivist learning theorists
• Core claim: When context systems reliably extend human cognitive processes — storing and retrieving information the human cannot hold in working memory — they become part of the cognitive system itself, not merely tools used by it.
• Key arguments:
  • Otto’s notebook (Clark & Chalmers): a notebook consulted reliably plays the same role as long-term memory
  • External systems that are always available, always consulted, and always trusted constitute genuine extended cognition
  • The coupling between human and context system is more important than the location of the storage
• Relation to site tenets: Deepens the Context as Infrastructure tenet philosophically — context infrastructure is not just data plumbing but an extension of the practitioner’s mind. Supports Symbiotic Intelligence (genuine augmentation, not replacement).

Position 3: Context as Organisational Culture and Practice (Sociotechnical View)

• Proponents: Workday (“context as culture”), Mastercard (“context by design”), knowledge management literature
• Core claim: Sustainable context infrastructure requires organisational practices — shared vocabulary, governance rituals, cross-team ownership — not just technical systems. Context must be “a living system” that evolves with the organisation’s understanding.
• Key arguments:
  • Shared language among humans (glossaries, domain ontologies) is the same resource AI systems use — AI readiness depends on human semantic alignment first
  • Context drift is an organisational failure before it is a technical one
  • Context ownership (who maintains it, who can modify it) determines whether AI behaves consistently
• Relation to site tenets: Strongly aligned with Pluralism of Perspectives — organisational context must include minority views, not just dominant definitions. Human Intent First — context is ultimately a representation of what humans in the organisation care about.

Position 4: Context as Labour (Critical / Resistance View)

• Proponents: Implicit in “oversight tax” literature; labour studies of AI-mediated work
• Core claim: Building and maintaining context infrastructure is invisible labour. The “translation tax” (Ja’dan Johnson) and “context amnesia” are not neutral technical limitations — they shift cognitive burden onto users who must prepare, maintain, and repair context in addition to doing their actual work.
• Key arguments:
  • Context maintenance is ongoing, uncredited overhead that accumulates across projects
  • Power asymmetry: organisations that have well-funded context infrastructure collaborate with AI more effectively, widening capability gaps
  • The cognitive overhead of context management can offset or exceed the efficiency gains from AI assistance
• Relation to site tenets: Tensions with Always Scalable (which assumes effort in = results out — context overhead may break this equation for under-resourced practitioners). Connects to the Oversight Tax article already in the system.

Key Debates

Debate 1: Whose job is context engineering?

• Sides: Data engineering teams (context is data infrastructure) vs. AI application teams (context is model configuration) vs. domain experts and practitioners (context encodes professional knowledge)
• Core disagreement: Who owns the context layer determines what gets encoded in it, whose knowledge is treated as authoritative, and who bears the maintenance burden
• Current state: Ongoing in 2026. Atlan has published a dedicated article on “Who Should Own the Context Layer: Data Teams vs. AI Teams?” — no settled answer

Debate 2: How much context is too much?

• Sides: “Bigger is better” (longer context windows eliminate the need for context engineering) vs. “Context rot is real” (more tokens degrades quality; curation is always necessary)
• Core disagreement: Whether scaling context windows eventually eliminates the curation problem, or whether the fundamental attention limitations of transformers make selective curation permanently necessary
• Current state: Current evidence (needle-in-haystack benchmarks, Anthropic 2025) favours context rot as persistent phenomenon, but the severity diminishes with more capable models

Debate 3: Does context infrastructure reinforce dominant knowledge?

• Sides: Efficiency view (standardised context speeds collaboration) vs. pluralism critique (encoded context reflects whoever built the glossary and defines the ontology)
• Core disagreement: Whether context systems can be made genuinely pluralistic or whether the act of codification necessarily privileges some perspectives over others
• Current state: Under-explored in the literature; strong connection to the Pluralism of Perspectives and Epistemologies tenet

Debate 4: At what granularity should context be managed?

• Sides: Session-level (manage per-conversation state) vs. project-level (shared context across all sessions in a project) vs. organisational-level (enterprise knowledge graph)
• Core disagreement: Which granularity provides the best balance between relevance (specific to task) and reuse (applicable across tasks)
• Current state: Practitioners (Ja’dan Johnson) advocate starting with session/project level and scaling up; enterprises (Atlan, Mastercard) build organisation-level infrastructure

Historical Timeline

Potential Article Angles

1. Concept article: “Context as Infrastructure” — Define the tenet directly as a practitioner concept. Explain what it means to treat context as infrastructure in day-to-day AI collaboration: the three layers (library/persistent domain knowledge, conversation/session state, memory/cross-session notes), and what maintaining these layers looks like for a systemic designer or product owner. Strong alignment with Human Intent First (context preserves intent across sessions) and Symbiotic Intelligence (infrastructure enables genuine ongoing collaboration, not repeated cold starts).

2. Topic article: “Context Amnesia and the Translation Tax” — Examine the failure mode before the solution. What is actually lost when sessions reset? What cognitive work must be repeated? Frame through the lens of professional knowledge workers (designers, POs, strategists) and explore what makes the overhead disproportionate. Connect to the Oversight Tax article already in the system. Moderate tension with Always Scalable (context maintenance is a real cost).

3. Gaps/void article: “Who Owns the Context Layer?” — Explore the unresolved question of ownership in multi-person design and strategy teams. Who decides what goes into the shared context? Who updates it? What happens when context encodes one person’s understanding and excludes others? Strong connection to Pluralism of Perspectives tenet — the context layer is not neutral.

4. Counterargument article: “The Limits of Context Engineering” — Argue the other side: that context infrastructure cannot solve the fundamental problem of intent specification, that well-maintained context can generate false confidence, and that the overhead of context management may reproduce the problems it solves.

When writing article 1, use the extended mind theory as background (not the focus), front-load the practical definition, and connect to the tenet via the “Relation to Site Perspective” section. Follow named-anchor pattern for forward references to related concepts.

Gaps in Research

• Empirical studies on professional context practices: Little research exists on how systemic designers, UX strategists, or product owners currently maintain context across GenAI sessions in practice. Most literature is either technical (how LLMs handle context) or enterprise-scale (how organisations build context layers). The individual professional level is under-documented.
• Pluralism in context systems: Almost no literature examines whether context layers can encode genuinely multiple perspectives, or whether the codification process inherently flattens them. This is the most significant gap relative to site tenets.
• Context cost accounting: The “translation tax” concept lacks empirical backing. How much time/effort do knowledge workers actually spend maintaining context? Does it scale with project complexity in a predictable way?
• Failure modes specific to design work: The literature is heavily weighted toward software engineering (code context) and enterprise data management. Context patterns for creative, abductive, double-diamond work are absent.
• Philosophical depth: Extended mind literature predates LLMs. Richer philosophical treatment of what it means for AI context infrastructure to constitute distributed cognition — and whether the coupling is genuine or illusory — is needed.

Citations

1. Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19. https://www.alice.id.tue.nl/references/clark-chalmers-1998.pdf

2. Rajasekaran, P., Dixon, E., Ryan, C., & Hadfield, J. (2025, September 29). Effective context engineering for AI agents. Anthropic Engineering. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents

3. Centizen Nationwide. (2026, January 12). Beyond prompts: Why context engineering is the next big shift in AI. Medium. https://medium.com/@centizennationwide/beyond-prompts-why-context-engineering-is-the-next-big-shift-in-ai-eee131688e4c

4. Johnson, J. (2025, June 30). Mastering the context workflow: A practical guide to human-AI collaboration. Medium. https://jadanjohnson.medium.com/mastering-the-context-workflow-a-practical-guide-to-human-ai-collaboration-fa481722dba3

5. Winks, E. (2026, February 12). How to implement an enterprise context layer for AI: 2026 guide. Atlan. https://atlan.com/know/how-to-implement-enterprise-context-layer-for-ai/

6. Javeed, H. (2025, November 26). Escaping context amnesia: Practical strategies for long-running AI agents. https://hadijaveed.me/2025/11/26/escaping-context-amnesia-ai-agents/

7. Vaswani, A., et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. https://arxiv.org/abs/1706.03762

8. Mcclung, P. (2025). Context engineering: Sessions, memory. Medium. https://medium.com/@patmcc1979/context-engineering-sessions-memory-0505b825a273