May 1, 2026
Context Engineering Architecture 2.0
From service design to context design
/ Context Engineering · Architecture 2.0
From service design to context design
From service design to context design
Analogy → enterprise strategy
Part 1: 'LLM service' lens.
Part 2: Architecture 1.0 vs 2.0.
Part 3: signals 2025–2026.
Parts 4–5: flexibility cost and strategy.
Why 'just add an LLM service' is too narrow a lens
New coordination unit
Before: request, service, business logic.
Now: context, decisions, tools, state.
AWS: workflows shift logic to agents.
'Added AI' = context/tools/decisions.
Architecture 1.0 — deterministic systems, Architecture 2.0 — managing uncertainty
A new layer on top, not a replacement
Architecture 1.0
Request, service, logic, response.
Contracts, transactions, master data.
Architecture 2.0
Request, context, tools/model, memory.
Models, tools, state, orchestration, MCP.
7 trends already changing architecture
The two most underestimated shifts in AI-native architecture
Context → separate layer
From prompt engineering to context engineering.
Context rot: more tokens ≠ better.
Durable execution
Agents live for minutes, hours, weeks.
Background mode + state checkpoints.
Memory and capability surface become architecture
Memory — not a metaphor
LangGraph: session + cross-session memory.
Semantic, episodic, procedural memory types.
Tool layer = integration layer
MCP tool: name, schema, contract.
Discovery and invocation by context.
Quality observability + least privilege
Evaluation → production
Probabilistic behavior → evaluation system.
LangSmith: online eval on production traces.
Security: least privilege
OWASP: minimal tools and scopes.
Host controls permissions and boundaries.
Decision pipeline and new scalability pattern
Event → Decision pipeline
AWS: cognition-augmented workflows.
Events are interpreted, not only routed.
Scale in tokens
Anthropic: context window = working memory.
Compaction + prompt caching for cost/latency.
Each of the 7 trends creates a new class of risks
AI-native load-bearing elements
Uncontrolled context → recall/latency/cost.
Stateless runtime → lost work.
Over-privileged tools → prompt injection risk.
No online eval → blind production.
AI as a control plane over legacy, platform, and knowledge systems
Coordination layer over core systems
Systems of record (1.0)
Transactions, accounting, master data, identity, permissions.
Regulatory logic, billing.
AI control layer (2.0)
Context assembly, tool mediation, model routing.
Memory, evaluation, approvals, durable orchestration.
Architecture 1.0 and 2.0 enrich each other
1.0 → 2.0: identity, audit, policy, rollback.
2.0 → 1.0: context, evals, HITL, routing.
AI-native layer integrates guardrails.
From fixed paths to context
Design unit: context, decisions, execution.
Architecture 2.0 manages uncertainty over 1.0.
Enterprise: AI = control plane over legacy.
tellmeabout.tech · ai4sdlc-research.space
Agentic architecture, context, security
Anthropic: Building effective agents.
MCP and tool integration.
OpenAI Agents SDK and Responses API.
LangGraph agents + OWASP LLM Top 10.
Architecture 1.0 → 2.0
Materials and links are in the "Book Cube" channel
Alexander Polomodov, Technical Director & Fellow, T-Technologies