concepts · article
Capability-Driven Self-Evolution of Agent Memory (PrisMem)
Oct 11
PrisMem is a method for evolving agent memory systems by optimizing individual capability dimensions rather than overall performance alone, avoiding the pitfall of holistic evolution obscuring capability-specific gains and regressions. It uses dependency-aware capability selection, history-guided diagnosis, and trace-guided integration to merge complementary improvements into a unified memory program, outperforming baselines by 10.54 and 7.83 percentage points on BEAM-1M and LongMemEval-M respectively.
concepts · article
Shared Company Agent vs. Personal AI Agents
Laura Entis · Oct 11
Every Inc. transitioned from employees running individual OpenClaw bots to a single shared company agent, since personal bots proved hard to maintain and track, while a shared agent accumulates institutional knowledge and improves as everyone interacts with it. Willie Williams argues the future splits agent use between one company agent at work and personal agents at home for relationship-based, non-transactional tasks.
concepts · article
Context Libraries for AI Agent Management
Oct 11
A context library is a collection of institutional and procedural knowledge stored in markdown files or external sources (Notion, Slack, etc.) that AI agents reference and update. Even expert Claude power users struggle with context engineering, improvising bespoke processes without guidance from AI systems themselves, leading to stale or lost context files.
concepts · article
Processing failed
Oct 11
Could not process content automatically.
concepts · article
Boundary Action Auditing Against Staged Prompt Injection
Oct 11
This paper identifies 'staged prompt injection,' where indirect prompt injection attacks exploit long-horizon AI agents across multiple causally connected steps to alter consequential actions. The authors propose 'boundary action auditing'—evaluating each pending tool call or message immediately before execution—and introduce Path-Aligned Attribution (PAA), a training-free auditor that traces the origin of each element in a pending action to detect attacker influence, achieving 86% block recall at low false-block rates on Claude Sonnet 5, outperforming existing auditors like ARGUS and VIGIL.
concepts · article
Self-Correcting Memory via Claim-Evidence Linking in OpenWiki
Sep 20
OpenWiki addresses memory staleness by linking each factual claim in its wiki to specific versioned evidence in source code, rather than just storing text. When the underlying evidence changes version, the claim is flagged as stale and remains uncertain until re-verified against current code, at which point the wiki is updated to stay in sync with the true source.
concepts · article
AI Agent Prompt Injection as CI/CD Supply Chain Threat
Aug 24
AI coding agents like Claude Code, Gemini CLI, and GitHub Copilot Agent integrated into CI/CD pipelines process untrusted GitHub metadata (issue titles, PR bodies, comments) as trusted prompt content, enabling attackers to inject malicious instructions without needing write access. This flaw class was exploited in real incidents like the Clinejection attack on Cline's npm package, and Anthropic's own Claude Code Action had a critical authorization bypass allowing unauthenticated actors to trigger repository compromise.
concepts · article
Processing failed
Aug 24
Could not process content automatically.
concepts · article
Claude Managed Agents: Decoupling Brain, Hands, and Session
Aug 1
Anthropic's Managed Agents service virtualizes agent components—session (event log), harness (the Claude loop), and sandbox (execution environment)—into separate interfaces that can be independently swapped or replaced, similar to how OS abstractions like process and file decoupled software from hardware. This solves a 'pet vs. cattle' infrastructure problem where coupling everything into one container made sessions fragile and undebuggable, and baked in assumptions (like requiring code to live in the same container) that broke when customers needed different infrastructure setups.
concepts · article
Two-Agent Pattern for Long-Running Coding Agents
Aug 1
Long-running AI agents lose all memory between context window sessions, causing them to either attempt too much at once and fail mid-implementation, or falsely declare a project complete. The solution is a two-part system: an initializer agent that sets up project scaffolding (init script, progress log, initial commit), followed by coding agents that make incremental progress each session while leaving clean, documented state for the next session.
concepts · article
Loops are Replacing Prompts. Verification is About to Be Your Biggest Problem.
Arjun Iyer · Jun 21
Three eras of AI coding — prompt-driven, spec-driven, loop-driven — and why verification becomes the binding constraint as the human moves up a level each era. Loop economics: total cost = iterations-to-verified × cost-per-iteration.
concepts · article
AI Agent Security and Prompt Injection Vulnerabilities
Airia Team · Jun 14
Agentic AI systems face a fundamental security flaw where they cannot distinguish between instructions and data, leading to prompt injection attacks. The 'Lethal Trifecta' combines access to private data, exposure to untrusted tokens, and exfiltration vectors, enabling attacks like EchoLeak and GeminiJack that steal sensitive data through hidden instructions in emails and documents.
concepts · article
AI Agent Orchestration Patterns for Production
JobsByCulture · Jun 14
Multi-agent AI orchestration coordinates multiple agents to solve complex tasks that single agents cannot handle due to context limits, specialization needs, or parallelism requirements. The Sequential Chain pattern runs agents in fixed sequence where each agent's output feeds the next, ideal for tasks with natural linear progression like document processing pipelines.
concepts · article
Compound Engineering 8-Step Framework Evolution
Kieran Klaassen · Jun 14
Compound engineering has evolved from a 4-step loop (brainstorm → work → review → compound) to an 8-step process that adds ideation and planning at the front, and polishing at the end. The framework recognizes that AI handles the middle work phases well, but humans remain crucial for initial vision-setting and final quality assessment.
concepts · article
Agent Engineering Framework and Definition
Latent.Space · Jun 14
Explores the challenge of defining AI agents, presenting multiple perspectives from OpenAI's TRIM model (model + instructions + tools + runtime) to Lilian Weng's definition (LLM + memory + planning + tools). Proposes six core elements: LLMs with tools, encoded intent, LLM-driven control flow, multi-step planning, long-running memory, and goal-oriented behavior.
concepts · article
Enterprise AI Tool Cost Management Strategies
Simon Willison · Jun 7
Uber implemented $1,500 monthly spending caps per employee for AI coding tools like Cursor and Claude Code after exceeding their 2026 AI budget in four months. This represents approximately 11% of median engineer compensation and suggests companies are finding real value in AI tools despite needing cost controls.
concepts · article
How Coding Agents Work with LLMs
Simon Willison · Jun 7
Coding agents are software harnesses that extend LLMs with additional capabilities through invisible prompts and callable tools. LLMs work by completing text sequences using tokens (not words), with providers charging based on token usage. Chat templated prompts simulate conversations but require replaying entire conversation history each time, making longer conversations more expensive.
concepts · article
Trust Layer for AI-Generated Office Files
Nate · May 31
AI can generate polished-looking spreadsheets and presentations that contain hidden errors like broken formulas or wrong data sources. A systematic verification workflow with source tracking, assumption logging, and hostile review prevents shipping confident-looking but incorrect work.
concepts · article
AI-Native Company Operations and Workforce
Lenny Rachitsky · May 31
Every, a 15-person company, generates 7-figure revenue with 100% AI-written code across 5 products. They use specialized AI agents for different tasks and employ an 'AI operations lead' to maximize team productivity through AI tools.
concepts · article
Agent-Native Software Architecture Paradigm
Dan Shipper · May 31
A new software development approach where AI agents, rather than traditional code, form the core of applications. Features are defined as prompts describing desired outcomes rather than step-by-step instructions, making software more malleable and accessible to non-programmers.
concepts · article
AI-Native Business Model and Organizational Structure
Dan Shipper · May 31
Every operates as an AI-first company combining media, products, and consulting with just 15 employees generating $1.2M ARR. Their model involves living in the future with AI tools, documenting observations, building missing solutions, and teaching what works. Everyone is a generalist who uses AI for all tasks, blending traditional job roles.
concepts · article
Harness Engineering and Adversarial AI Architecture
Eric · May 31
Prompt engineering has reached its limits for complex autonomous tasks, leading to a new paradigm called 'Harness Engineering' that focuses on building structured environments around AI agents. Anthropic's breakthrough uses a GAN-inspired architecture with separate Generator and Evaluator agents creating adversarial feedback loops to overcome AI's inability to self-critique effectively.
concepts · article
KPMG-Anthropic Strategic AI Alliance for Enterprise
May 31
KPMG announced a global alliance with Anthropic to integrate Claude AI across its entire workforce of 276,000+ employees and core business operations. The partnership embeds Claude into KPMG's Digital Gateway platform for client work in tax and legal services, and establishes KPMG as Anthropic's preferred partner for private equity.
concepts · article
PwC Claude Enterprise AI Implementation Strategy
May 31
PwC and Anthropic expanded their partnership to deploy Claude across PwC's global workforce of hundreds of thousands, focusing on agentic technology builds, AI-native deal-making, and enterprise function reinvention. The collaboration includes training 30,000 professionals and launching a Claude-based finance business unit, with production deployments already cutting delivery times by up to 70%.
concepts · article
LLM Knowledge Bases for Business Intelligence
May 31
LLM knowledge bases transform raw organizational data into self-organizing, queryable intelligence systems that improve with use. Andrej Karpathy demonstrated this approach by having AI compile and maintain a 100-article knowledge base without complex RAG infrastructure. The AI-driven knowledge management market is projected to reach $11.24 billion by 2026.
concepts · article
Enterprise LLM Wiki Knowledge Management Pattern
May 31
Karpathy's LLM Wiki pattern uses an AI agent to automatically maintain a personal knowledge base by processing raw sources into structured wiki pages with cross-references and contradiction detection. The pattern works well personally but fails at enterprise scale because companies lack a dedicated curator and need automated ingestion from existing work tools rather than manual curation.
concepts · article
Forward Deployed Engineers in AI Companies
May 31
Major AI companies like Google, OpenAI, and Anthropic are rapidly hiring Forward Deployed Engineers (FDEs) through simplified processes and separate organizations. These roles involve integrating AI systems into enterprise customers' workflows and operations. The role is evolving from platform engineering toward solutions architecture and consulting.
concepts · article
Forward Deployed Engineers in Enterprise AI
May 31
AI vendors are embedding Forward Deployed Engineers (FDEs) to help enterprises implement agentic AI solutions, but Gartner predicts 70% of companies will abandon these projects by 2028 due to high costs and lack of internal capabilities. The key is ensuring knowledge transfer and capability building rather than creating vendor dependency.
concepts · article
Palantir's Forward Deployed Engineer Enterprise Model
MindStudio Team · May 31
Palantir's Forward Deployed Engineer (FDE) model embeds engineers directly inside client companies to build and deploy AI solutions, bridging the knowledge gap between AI capabilities and business requirements. This approach drove 640% returns and is now being adopted by Anthropic and OpenAI for enterprise deployment.
concepts · article
AI Agent Memory Benchmarks and Architectures 2026
May 31
Standardized benchmarks LoCoMo, LongMemEval, and BEAM now measure AI agent memory performance across dimensions like recall, temporal reasoning, and multi-session continuity. Top systems achieve 92.5 on LoCoMo and 94.4 on LongMemEval, with major gains in temporal reasoning (+29.6 points) and multi-hop queries (+23.1 points).
concepts · article
Claude Memory Architecture for Persistent Context
May 31
Agent failures in long-running workflows stem from context drift, not model limitations. The solution is engineered memory architecture using three layers: CLAUDE.md for human instructions, auto memory for model learnings, and managed stores for versioned persistence. This shifts focus from endless chat history to selective state reconstruction.
concepts · article
LLM-as-a-Judge for Automated Model Evaluation
Karyna Naminas · May 31
LLM-as-a-Judge uses powerful language models like GPT-4 to automatically evaluate AI outputs at scale, achieving 80% agreement with human evaluators while providing 500x-5000x cost savings over manual review. The approach involves prompting capable models to assess quality, safety, and relevance of other models' outputs based on specified criteria.
concepts · article
LLM as Judge Pattern for Agent Safety
MindStudio Team · May 31
A safety pattern where a second AI model evaluates and approves agent actions before execution, preventing costly mistakes in production workflows. The judge acts as a gatekeeper that can pause, revert, or route problematic actions to humans when rule-based guardrails aren't sufficient.
concepts · article
Multi-Agent AI Systems Architecture and Performance
May 17
Claude's Research feature uses multiple AI agents working in parallel to handle complex, open-ended research tasks that require dynamic planning and exploration. The multi-agent approach outperformed single-agent systems by 90.2% on internal evaluations, with token usage explaining 80% of performance variance.
concepts · article
Context Engineering vs Prompt Engineering
May 17
Context engineering is the evolution of prompt engineering, focusing on optimizing the entire set of tokens and information available to an LLM during inference, not just the prompts. As AI systems become more agent-like with multi-turn interactions, managing context becomes critical due to 'context rot' - the degradation of model performance as context length increases.
concepts · article
Real-time Search Quality Evaluation Systems
May 17
Sierra developed an evaluation system that measures AI agent search quality against real conversations by creating daily 'golden datasets' from anonymized customer interactions. The system uses retrieval metrics like recall, precision, and nDCG to identify search failures and drive continuous improvement, leading to up to 16 percentage point improvements in resolution rates.
concepts · article
Fine-tuning Agents with Reverse-Engineered Training Data
May 17
Shopify fine-tuned Qwen3-32B to generate workflow automations from natural language by reverse-engineering training data from existing user workflows. They achieved 2.2x speed improvement and 68% cost reduction compared to frontier models by starting with validated production workflows and working backwards to generate plausible user queries.
concepts · article
AI Agent Workflows for 10x Engineering Productivity
Rhea Purohit · May 17
Two engineers at Every ship like a team of 15 by designing AI agent workflows that compound - using a meta-prompt that writes prompts to transform rough ideas into detailed GitHub issues, then carefully planning before coding. They emphasize fixing problems early in the workflow when stakes are low, inspired by Andy Grove's High Output Management principles.
concepts · article
LLM Wiki vs RAG Knowledge Management
May 17
Andrej Karpathy's LLM wiki is a three-folder markdown system that loads structured knowledge directly into LLM context, while RAG retrieves chunks dynamically from vector stores. LLM wiki excels for personal-scale knowledge bases (up to 100 articles) with 95% token savings, while RAG scales to enterprise-level millions of documents.
concepts · article
Authorization Propagation in Multi-Agent AI Systems
May 17
Multi-agent AI systems face a distinct security challenge beyond prompt injection: maintaining authorization invariants as non-human agents delegate tasks, retrieve data, and synthesize results across changing boundaries. This 'authorization propagation' problem involves three sub-problems: transitive delegation, aggregation inference, and temporal validity that classical access control models don't fully address.
concepts · article
Agent-Native Software Development Principles
Jan 13
Five core principles for building software where features are outcomes described in prompts rather than written code. Agents use atomic tools in loops to achieve objectives, with capabilities emerging from tool composition rather than explicit programming.
concepts · article
Knowledge Distribution with Claude Skills System
Hedgineer Technologies · Jan 13
Hedgineer Technologies systematized Anthropic's Claude Skills to automatically distribute institutional knowledge across four technical domains (AI, Data, Infrastructure, UI). Skills are model-invoked rather than user-invoked, meaning Claude automatically applies relevant expertise based on context without requiring engineers to know which skills exist.
concepts · article
LLM Agent Architecture Patterns and Design
Jan 12
Anthropic distinguishes between workflows (LLMs orchestrated through predefined code paths) and agents (LLMs that dynamically direct their own processes). The most successful implementations use simple, composable patterns rather than complex frameworks, building on augmented LLMs as the foundational block.