34:14FULL TALK May 2026
Harness Engineering Deep Dive
The full introduction to engineering coding agents: model choice, retrieval, memory, tool use, feedback loops, and system architecture.
Harness Engineering Videos
34:14Short explainers on tools, memory, routing, verification, and agent teams. Pick a topic or find a specific idea.
19 shorts 7 topics by Rajiv Shah19 videos
SHORT1:056 ways to build a better agent
A cinematic guide to harness choice, tools, context, model routing, goals, and when to bring in an agent team.
SHORT0:34October cybersecurity threats are overwhelming
A satirical look at security updates, a compromised dependency, and deployment disruptions.
SHORT1:32System prompts encode the workflow
System prompts carry workflow decisions: when to plan, retry a tool, edit code, and run tests.
SHORT1:39Tool schemas and harness overhead
Tool definitions consume context too. A simpler harness can change the cost of the same model.
SHORT1:28Dynamic workflows
Give an agent a tool and let it write orchestration code, delegate work, and bring the results together.
SHORT1:10CupidBot: retrieval methods
CupidBot explains Find, BM25, embeddings, rerankers, GraphRAG, and agentic search through one urgent question.
SHORT1:28CliffCompaction
Why deliberately trimming old tool output can help an agent continue a long coding task.
SHORT1:05Hermes: memory in layers
Build agent memory in layers, from repository rules and session history to reusable and self-written skills.
SHORT1:20FrontierMemory: lessons from work logs
Turn past work logs into useful lessons, instructions, and skills for the next task.
SHORT1:28Recursive Language Models
Let the model explore a large context with code instead of placing every detail in its prompt.
SHORT1:46Four ways to route models
Rules, classifiers, cost-aware selection, and cascades offer four ways to choose the next model.
SHORT1:19Verifiers’ Law
Why checking an answer can be easier than finding it, and how that shapes progress in coding and math.
SHORT1:54Harness evolution or just more retries?
Test a self-improving harness against simple retries at the same budget and on held-out tasks.
SHORT2:21Long-running agents overfit validation
A competition story about agents optimizing validation noise, with four ways to guard against it.
SHORT1:27Multi-agent coding needs isolation
Managers, dependency graphs, worktrees, merges, and tests make parallel coding work reviewable.
SHORT1:12Office politics at machine speed
An agent swarm finds more bugs, burns more tokens, and sometimes sabotages its own teammates.
SHORT1:49NeuroGolf: managing agent work
What a coding competition teaches about task queues, memory, isolated workspaces, and approval gates.
SHORT1:39More agents is not a strategy
Tool overload, coordination costs, and amplified errors explain why adding agents can hurt performance.
SHORT1:01Patterns and problems in multi-agent systems
A short on Anthropic’s research into patterns and problems in emerging multiagent systems.
More AI videos: YouTube @rajistics_ai · Instagram @rajistics_ai.
Longer walkthroughs
The Dynamic Workflows deep dive pairs with P08: Dynamic Workflows. It follows the model as it writes orchestration code and delegates work.
Ready to try a harness decision yourself? Explore the projects or start the workshop.
For papers, code, and the sources behind the talk, explore References and Further Reading.