
Part 1: The Gods, Interns, and Cogs Framework
- The Three AI Use Cases: Breunig groups AI applications into Gods, Interns, and Cogs to make very different product ambitions easier to discuss. — The 3 AI Use Cases.
- Gods as Human-Replacement Systems: In his taxonomy, Gods are autonomous, general systems aimed at replacing human work—not today's supervised copilots. — The 3 AI Use Cases.
- The Barrier to Building Gods: Breunig expects God-like systems to require breakthroughs, many researchers, and at least billions of dollars, limiting who can pursue them. — The 3 AI Use Cases.
- Interns as Expert Copilots: Interns are supervised copilots that help domain experts with grunt work such as documentation, details, and ideation. — The 3 AI Use Cases.
- Expert Oversight: An Intern's output is reviewed by an expert, who remains responsible for catching mistakes before the work goes further. — The 3 AI Use Cases.
- Error Tolerance for Interns: Because an expert checks the output, Intern-style tools can tolerate more mistakes than unsupervised AI components. — The 3 AI Use Cases.
- The Value of Interns: In his October 2024 assessment, Intern-style copilots were delivering much of AI's realized value by extending expert workers. — The 3 AI Use Cases.
- Cogs as Focused Functions: Cogs are AI components built to perform one narrow task well inside a pipeline or interface. — The 3 AI Use Cases.
- Cogs Need Reliability: Cogs often run with little expert supervision, so Breunig sets a low error tolerance for them. — The 3 AI Use Cases.
- A Taxonomy for Bottlenecks: Breunig uses the three buckets to clarify questions about AI products and identify each category's constraints. — The 3 AI Use Cases.
Part 2: Context Engineering Fundamentals
- What Context Engineering Manages: Breunig describes context engineering as systematically choosing the information and tools an application supplies to a model for a particular outcome. — Why Context Engineering Matters.
- Manage the Working Context: Large context windows do not make every document or tool useful; Breunig argues for selecting and maintaining what a model actually needs. — How to Fix Your Context.
- Context Is Assembled: His Claude Code analysis shows an agent's context assembled dynamically from instructions, tools, user material, and conditional application state. — How Claude Code Builds a System Prompt.
- More Than the Task Prompt: His Claude Code case study shows agent behavior depends on tools, application instructions, and session information beyond a short task request. — How Claude Code Builds a System Prompt.
- Structure Context for Maintenance: Breunig suggests structuring context so an application can assemble it, prune irrelevant material, and preserve core instructions. — How to Fix Your Context.
- Prompting and Context Are Different: He distinguishes short conversational prompting from designing context for applications and agents; neither term is inherently superior. — Prompts vs. Context.
- Inputs Influence Outputs: Breunig warns that irrelevant or erroneous context can influence a model's response, so context quality matters. — How to Fix Your Context.
- Supply Relevant Information: Selective retrieval can add information relevant to a task instead of loading every available document into context. — How to Fix Your Context.
- Agents Accumulate Context: Multi-step agents gather documents, tool outputs, and history, making active context management especially important. — How Long Contexts Fail.
- Naming a Shared Practice: Breunig argues that context engineering gave practitioners common language for work they were already doing; he credits Karpathy with popularizing it. — Why Context Engineering Matters.
Part 3: The Mechanics of Context Rot
- Context Poisoning: An error or hallucination in an agent's context can be repeatedly reused, causing later reasoning to build on a false premise. — How Long Contexts Fail.
- Context Distraction: In some long agent runs, accumulated history can distract a model from developing a new plan and make it repeat earlier actions. — How Long Contexts Fail.
- Context Confusion: Irrelevant documents or excessive tool descriptions can affect tool use and lead to lower-quality responses. — How Long Contexts Fail.
- Context Clash: Agent context can contain conflicting information from messages, documents, and tools, complicating later decisions. — How Long Contexts Fail.
- Relevant Facts Can Be Missed in the Middle: In the tasks tested by Liu and colleagues, models often used relevant information less effectively when it appeared in the middle of a long input than at the beginning or end. — Lost in the Middle Study.
- How Long Contexts Fail: Breunig argues that expanding a context window can introduce poisoning, distraction, confusion, and clash rather than reliably improving answers. — How Long Contexts Fail.
- Long Inputs Can Displace Instructions: Breunig notes a Databricks finding that some models summarized long supplied material while neglecting the task instructions. — How Long Contexts Fail.
- The Attention Cost of Extra Context: Breunig cautions that even irrelevant tool definitions or documents can influence an agent's response, so adding context has a cost. — How Long Contexts Fail.
Part 4: AI Product Strategy and Subsumption
- The Subsumption Window: Breunig calls the period between a product launch and a future model's ability to reproduce its core function the subsumption window. — Will the Model Eat Your Stack?.
- Build Around the Model: Breunig advises AI product teams to focus on one or two parts of their stack that may resist model improvements, such as UI or data, then reassess as models change. — Will the Model Eat Your Stack?.
- Specs and Tests Can Guide Agent Coding: In his whenwords experiment, Breunig used a detailed text specification and conformance tests to let agents generate implementations in different languages. — The Rise of Spec Driven Development.
- Keep Spec, Tests, and Code in Sync: Breunig's spec-driven-development triangle treats implementation as feedback: code reveals decisions that should update both the specification and its tests. — The Spec-Driven Development Triangle.
- Reassess as Models Improve: After finding traction, Breunig recommends checking whether model advances are absorbing an application's advantage and moving to the next hard problem if needed. — Will the Model Eat Your Stack?.
- UI Can Differentiate AI Products: Comparing AI search products in 2024, Breunig argued that a cleaner user experience can matter more to users than model differences. — AI Search Apps and UI.
- Data Can Extend a Product's Window: He identifies access to useful data as one possible stack advantage when out-of-the-box models cannot solve a product's task well. — Will the Model Eat Your Stack?.
- A Communication Tax in New Fields: Breunig argues that unsettled terminology makes emerging AI fields harder to discuss, while shared categories reduce the cost of productive exchange. — The 3 AI Use Cases.
Part 5: Location Intelligence and Geospatial Standards
- Make Place Easier to Use: Breunig compares a potential shared standard for places to the way common time standards let software exchange timestamps without custom translation. — Towards Standardizing Place.
- Standardize Places, Not Just Coordinates: Coordinates standardize where a point is; Breunig argues that useful place data also needs shared identities for roads, buildings, stores, and other human-defined entities. — Towards Standardizing Place.
- The Join Problem: Without a common place identifier, organizations must repeatedly match records that describe the same real-world feature in different datasets. — Towards Standardizing Place.
- Placekey as a Join Key: Placekey assigns shared identifiers to physical places so compatible datasets can be matched and joined more easily. — Join Data with Placekey.
- GERS Identifies Overture Features: Overture's GERS assigns stable identifiers to eligible features in its reference map, allowing matched datasets to join by ID. — What Is GERS?.
- Analysis Beyond Map Views: Breunig argues that shared place IDs could let more people analyze location data through familiar joins and tables, even when no map is needed. — Towards Standardizing Place.
- A Shared Identifier Across Datasets: A GERS ID lets publishers and users refer to the same Overture feature across datasets, reducing repeated feature-by-feature conflation after an initial match. — What Is GERS?.
Part 6: Data Integrity and Information Environments
- Model Weights and Context Both Matter: Applications must account for both what a model learned during training and the material placed in its context; retrieval alone cannot reliably override a model's strong prior. — Don't Fight the Weights.
- Reliability Before Broad Agent Adoption: Breunig argues that unreliable internally built agents struggle to gain enterprise adoption, so teams should begin with constrained, testable tasks. — Enterprise Agents Have a Reliability Problem.
- The Harness Shapes an Agent's Context: An agent harness decides what enters the model's context, how the loop runs, and what persists across sessions; its surrounding environment matters to behavior. — Harnesses Are Situated Agents.
- Cache Boundaries in Agent Prompts: In his analysis of Claude Code, Breunig identifies a cache-boundary marker separating globally reusable system-prompt material from session-specific content. — How Claude Code Builds a System Prompt.
- Prune Irrelevant Context: Breunig recommends pruning stale or irrelevant context so an agent works from a smaller, more useful information set. — How to Fix Your Context.
Part 7: Engineering Reliability and Tool Harnesses
- Agents Are More Than a Model: Breunig describes agents as a core loop situated within a harness that may manage sessions, tools, environments, memory, skills, teams, and organizational policies. — Harnesses Are Situated Agents.
- What the Agent Harness Controls: A harness determines what reaches an agent's context, how its loop runs, and what information gets saved. — Harnesses Are Situated Agents.
- Engineering Still Matters When Agents Code: Breunig recommends end-to-end tests, documented intent, current specifications, and attention to maintenance and security even when agents make code cheap. — 10 Lessons for Agentic Coding.
- Build an Eval Before Choosing a Model: A task-specific evaluation lets builders compare models and iterate on prompts or pipelines against the product goal instead of relying on a default model. — Your Eval Is More Important Than the Model.
- Choose a Relevant Tool Loadout: Select only tool definitions relevant to a task; as tool menus expand, overlapping descriptions can confuse models and hurt selection. — How to Fix Your Context.
- Quarantine Independent Contexts: Split suitable subtasks into separate threads so each agent receives a shorter, more relevant context; tasks requiring shared context may not benefit. — How to Fix Your Context.
Part 8: The Cultural and Linguistic Shift of AI
- Shared Language Shapes a Field: Breunig argues that a shared term can make an emerging practice easier to discuss and develop collectively; he uses “context engineering” as his example. — Why “Context Engineering” Matters.
- Why “Context Engineering” Caught On: Breunig credits Andrej Karpathy with popularizing “context engineering” and argues that the phrase captured work practitioners were already doing. — Why “Context Engineering” Matters.
- Security as a Token-Budget Contest: Breunig suggests AI-assisted cybersecurity could become a contest of token budgets: defenders may need to spend more finding and fixing exploits than attackers spend exploiting them. — Cybersecurity Looks Like Proof of Work Now.
- Three Models of Software Creation: Breunig contrasts the planned, closed Cathedral and open, collaborative Bazaar with a third AI-era pattern: idiosyncratic, sprawling personal software he calls the Winchester Mystery House. — The Cathedral, the Bazaar, and the Winchester Mystery House.
- The Limits of Cheap Code: Breunig warns that coding agents make implementation cheap while review, coordination, and feedback remain costly; personal 'Mystery Houses' can grow sprawling without shared conventions. — The Cathedral, the Bazaar, and the Winchester Mystery House.