Complete 21-part series

The Inference and Automation Field Guide

From one inference request to durable loops, reinforcement learning, swarms, and the economics of verified automation. Read in order, or enter at the decision you are facing.

1. 01Three Factories. One Missing Layer.A precise map of compute factories, model factories, software factories, and the independent proof boundary above them.17 min →
2. 02From Prompt to ProofWhy first-token latency is not automation latency, and what an end-to-end inference trace must prove.15 min →
3. 03The Token Bill LiesA reproducible model for the costs that appear after the token invoice ends.14 min →
4. 04Capacity Is a QueueA queueing model for inference utilization, bursts, agent fan-out, memory admission, and tail-latency goodput.12 min →
5. 05The Cache LadderA correctness-first guide to answer reuse, prefix reuse, KV state, invalidation, and cache economics.13 min →
6. 06The Batch Never WaitsIteration-level scheduling, KV admission, prefill interference, fairness, and goodput in one request trace.12 min →
7. 07The Draft Is Allowed to Be WrongA lossless draft-and-verify mechanism, breakeven model, and workload test for faster autoregressive decoding.13 min →
8. 08Four Bits. Full Consequences.A tensor-by-tensor guide to bits, calibration, kernels, task quality, serving capacity, and outcome economics.13 min →
9. 09Route the WorkA constrained policy for model selection, cascades, calibrated confidence, fallback, and cost per verified outcome.13 min →
10. 10The Window Is Not the MemoryA token-allocation system for policy, evidence, tools, history, output, and verified long-context value.13 min →
11. 11Seven Places RAG Can LieA stage-by-stage failure map for corpus coverage, retrieval, reranking, generation, citations, latency, and supported outcomes.13 min →
12. 12Make Invalid Tokens ImpossibleA grammar-to-receipt architecture for structured generation, semantic validation, authorization, and safe tools.13 min →
13. 13The Loop Is the RuntimeThe minimum state, transition, authority, evidence, budget, recovery, and stop semantics of a dependable agent.14 min →
14. 14TokenMaxingA marginal-value framework for retries, parallel samples, search, revision, context replay, and verification.13 min →
15. 15After the DemoWho owns state when agents own the work? Durable execution, idempotency, recovery, compensation, and replay for AI-native development.13 min →
16. 16No Trace. No Truth.The outcome trace for models, context, tools, state, policy, artifacts, cost, and proof.12 min →
17. 17The Benchmark Ends Too SoonWhy model scores stop before review, rework, integration, deployment, maintenance, and ROI begin.13 min →
18. 18Reward Is a Bug ReportHow environments, actions, verifiers, resets, costs, and shortcuts train multi-hop agents.12 min →
19. 19The Last Step Gets Too Much CreditWho deserves reward across delayed tools, rejected branches, verifiers, and agent teams?13 min →
20. 20Swarms Multiply ErrorsWhen parallel agents create real diversity, and when they create correlated failure and coordination debt.13 min →
21. 21The Last SeatThe rise and fall of software GCCs and ODCs, and the end of the seat as the unit of capacity.12 min →