← Latest brief

Brief #181

25 articles analyzed

Context engineering is fragmenting into two competing philosophies: protocol-based standardization (MCP) versus demonstration-based specification (Skills/Codex). Practitioners are discovering that the infrastructure bottleneck isn't model capability—it's the gap between how humans demonstrate workflows and how AI systems preserve that context across sessions.

Demonstration Beats Documentation for Implicit Workflows

EXTENDS prompt-engineering — moves beyond static prompt templates to dynamic workflow capture

Practitioners are replacing written documentation with recorded demonstrations for complex workflows because autoregressive models learn better from sequential action patterns than from text descriptions of implicit preferences. This inverts traditional knowledge capture: show the work, let the model extract the rules.

Record one complex workflow as a demonstration with explicit parameter declarations before attempting to document it in text. Compare which format produces more accurate agent execution.
@shao__meng: 有些工作流很难用纯文字精准描述...

Demo-as-spec pattern: workflows with implicit preferences (expense reports, issue config) are captured via demonstration + semantic extraction rather than written documentation. Codex Skills use parameterization to make demos reusable.

@doodlestein: The more I think about how digital twin models...

Scientific method applied to agent workflows: hypothesis registration before action, structured result gathering. Workflow must be 'done once' to establish clear methodology, then agents can replicate.

How to use Agentic Coding Tools like Claude Code or Codes Effectively

Context management through explicit file curation and instruction clarity improves Claude Code effectiveness. Breaking tasks into clear subproblems with explicit roles (writing vs reviewing) mirrors demonstration-based structuring.


MCP Context Poisoning is the New Supply Chain Attack

EXTENDS model-context-protocol — existing graph treats MCP as integration protocol, this reveals security architecture implications

As MCP servers become context bridges with tool execution privileges, they represent high-value attack surfaces. Compromising a single MCP server poisons every downstream agent's context, making server authentication and context validation critical infrastructure concerns.

Audit every MCP server for credential scope, implement context source validation, and enforce time-limited permissions for agent tool access. Treat MCP servers as critical infrastructure with same security rigor as API gateways.
Enterprise-Grade Security for the Model Context Protocol (MCP): Frameworks and Mitigation Strategies

Identifies MCP servers as trust boundaries: poisoned servers compromise entire context chains. Zero-trust architecture required with server identity verification, data source validation, least-privilege permissions, and context access auditing.

Multi-Agent Coherence Requires Central Knowledge, Not Just Orchestration

EXTENDS multi-agent-orchestration — existing graph focuses on communication patterns, this identifies shared state as critical missing piece

Teams deploying multi-agent systems are discovering that agent-to-agent communication alone creates context drift. Coherent output requires a central knowledge base that all agents read from and write to, plus supervisory reconciliation to prevent inconsistency compounding across iterations.

If running multiple agents, implement a shared knowledge base (vector store or graph database) that all agents can query and update, plus a supervisory agent that reconciles conflicting outputs before final execution.
Multi-Agent Architecture for LLM Applications | Aim Reply

Pattern: central knowledge repository plus supervisory reconciliation layer. Agents maintain local context but global consistency requires centralized shared state and arbitration. Without it, agents drift into incoherence.

Session State Checkpointing is Framework Selection Criterion

EXTENDS state-management — existing graph treats state as implementation detail, this elevates it to primary selection criterion

Practitioners are choosing agent frameworks based on whether they checkpoint conversation state throughout sessions, not just on speed or model compatibility. The ability to preserve and resume context mid-conversation has become more valuable than execution speed.

When evaluating agent frameworks, prioritize session checkpointing and state resumption over raw execution speed. Test whether the framework can recover mid-conversation after interruption without context loss.
Stop wasting time choosing the wrong AI framework! LangGraph vs. CrewAI✨

LangGraph's key differentiator is checkpointing session state throughout conversations (persistence/context preservation) while CrewAI optimizes for speed. Framework selection now hinges on stateful controlled systems versus autonomous fast systems.

Context Accessibility Gap Blocks AI Adoption More Than Capability

EXTENDS tool-integration-patterns — existing graph focuses on technical integration, this identifies adoption barrier as context flow interruption

Non-technical users avoid AI tools not because models lack capability, but because integrating AI into actual workflows requires developer-level setup. The human becomes the manual context bridge, doing copy-paste integration between AI and work tools. Distribution accessibility is the actual competitive moat.

Measure how many manual steps users take to move context from AI output into their actual work tools. Each step is an adoption barrier. Prioritize embedding AI directly in work interfaces over building separate AI tools.
@PereraBinoy: What is AI for 99% of the world? It's a frontier model like GPT sitting behin...

Integration Tax Pattern: easier non-developer access to intelligence without context setup equals wider adoption. Every API key, config file, or developer requirement creates context accessibility barrier. AI isolated in tabs means intelligence resets each session.

Context Inference Mismatch When Multiple Problem Frames Available

When AI has access to multiple versions of context (public repo plus internal code, generic task plus domain-specific knowledge), it may select a different problem framing than the user intended. Managing what context is available matters as much as how you phrase requests.

Before running AI requests, audit what context sources the model can access. Explicitly exclude or deprioritize context that might trigger unintended problem framings. Test whether removing context improves output accuracy.
@RhysSullivan: asked codex to check it's source code and it started to reverse engineer the...

Model had access to desktop app source code (not just public repo). Request interpreted as reverse engineering task rather than code review because additional context reframed the problem. Clarity requires clear problem scoping given available context.