# Can't believe no one talks about this... Source: https://www.youtube.com/watch?v=kJPvfoLtFFY Recap page: https://rapidrecap.app/video/kJPvfoLtFFY Generated: 2026-03-05T11:03:39.448+00:00 --- ## Quick Overview The shift from 'Context Engineering' (prompting, context window management) to 'Harnesses Engineering' (across-session context management, workflows, tooling, and verification loops) is crucial for enabling long-running, autonomous AI agents that can handle complex tasks reliably by December 2025. **Key Points:** - The progression in AI agent development moves from Context Engineering (focusing on single-session prompts and context windows) to Harnesses Engineering, which manages state across sessions. - Key components of effective harnesses include explicit cross-cutting boundaries in domain architecture, such as enforcing invariants via custom linters and structural tests, not just documentation. - Two primary failure modes for long-running agents were identified: agents attempting one-shot solutions and running out of context, and agents prematurely declaring jobs done without proper end-to-end verification. - Anthropic's solution involves a two-part structure: an Initializer agent to set up the environment (using init.sh, progress files, and git commits) and a Coding agent for incremental progress. - Vercel's experience with their d0 text-to-SQL agent showed that stripping specialized tools down to generic ones (like bash commands: grep, cat, ls) improved performance by 3.5x faster execution, 37% fewer tokens, and a 100% success rate. - Effective agent workflows require rigorous verification (build -> verify -> fix until evidence matches spec) and an initial environment setup that provides all necessary context, including a feature list and clean state. - The goal is Agent Legibility, where agents can reason directly from the repository itself, using structured documentation (like AGENTS.md) as a system of record, rather than relying on unseen knowledge. ![Screenshot at 00:33: Slide detailing the core shift from Context Engineering \(prompt & context window focus\) to Harnesses Engineering \(across-session context mgmt, workflow, tooling, verification loops\).](https://ss.rapidrecap.app/screens/kJPvfoLtFFY/00-00-33.jpg) **Context:** This presentation discusses the evolution of building reliable, long-running autonomous AI agents, contrasting the older method of 'Context Engineering' with the newer paradigm of 'Harnesses Engineering.' The speaker references recent industry experiments and publications from Anthropic (Claude Agent SDK) and Vercel (d0 SQL agent, Cursor browser building) to illustrate the challenges of context management and the solutions developed to achieve greater agent autonomy and reliability. ## Detailed Analysis The video argues that the state of AI agent development is undergoing a major shift, moving from simple 'Context Engineering' to robust 'Harnesses Engineering.' Context Engineering focused primarily on optimizing prompts and managing the immediate context window, which is insufficient for long-running, complex tasks. Harnesses Engineering addresses this by incorporating across-session context management, defined workflows, specialized tooling, and crucial verification loops. The speaker highlights that the evolution is necessary because LLMs are becoming capable enough to handle complex, multi-stage tasks spanning hours or days, exemplified by the December 2025 'WFT' event prediction. Key industry examples, like Anthropic's Claude Agent SDK, identified failure modes where agents either tried to do too much at once ('one-shotting') or prematurely declared tasks complete without proper testing. The solution proposed involves a two-part system: an Initializer agent to set up a clean, documented environment (using init.sh and progress files) and a Coding agent that makes incremental progress, ensuring the environment is left in a 'clean state' after each step. Furthermore, Vercel's success in simplifying their text-to-SQL agent (d0) by replacing specialized tools with generic bash commands (grep, cat, ls) resulted in a 3.5x speedup, 37% token reduction, and 100% success rate, reinforcing the principle that generic, well-understood tools are often superior to brittle, specialized ones. The ultimate goal is 'Agent Legibility,' where the agent can reason about the entire codebase by treating the repository's structure (like AGENTS.md) as its system of record, enforced by custom linters and structural tests. ### The Paradigm Shift - Context Engineering vs. Harnesses Engineering: Context Engineering focused on prompt & context window - Harnesses Engineering focuses on across-session context mgmt, Workflow, Tooling, and Verification loops ### Key Learnings from Long-Running Agents (Anthropic/Claude) - Agents fail due to one-shotting (running out of context) or premature job completion without verification - Solution involves an Initializer Agent (setup) and Coding Agent (incremental progress with clean state commits) ### Vercel's Simplification Success - Specialized tools were replaced by generic bash commands (grep, cat, ls) - Results: 3.5x faster execution, 37% fewer tokens, 100% success rate (up from 80%) ### Verification is Mandatory - Agents fail when they stop at 'looks right' - Harnesses must enforce a build -> verify -> fix loop until evidence matches the spec ### Architecture and Legibility - Documentation alone is insufficient; constraints (enforced via linters/tests) are prerequisites for speed - Repository knowledge (code, docs, schemas) must be the system of record, accessible contextually to the agent ![Screenshot at 00:33: Slide showing the evolution from Context Engineering to Harnesses Engineering, emphasizing the need for cross-session context management and verification loops.](https://ss.rapidrecap.app/screens/kJPvfoLtFFY/00-00-33.jpg) ![Screenshot at 00:46: Diagram illustrating the 2023 AutoGPT architecture, showing multiple agents interacting with a Task Queue, Memory, and the User.](https://ss.rapidrecap.app/screens/kJPvfoLtFFY/00-00-46.jpg) ![Screenshot at 01:03: Windows PowerShell terminal showing the setup process for an Auto-GPT agent with defined goals.](https://ss.rapidrecap.app/screens/kJPvfoLtFFY/00-01-03.jpg) ![Screenshot at 01:32: Diagram of the 'Ralph Loop' for Deep Agents: Task -\> DeepAgent -\> Filesystem -\> Loop to keep agent running.](https://ss.rapidrecap.app/screens/kJPvfoLtFFY/00-01-32.jpg) ![Screenshot at 02:46: Slide summarizing the three requirements for long-running autonomous agents: Memory Context + Cron/Task list + Full computer access.](https://ss.rapidrecap.app/screens/kJPvfoLtFFY/00-02-46.jpg)