StewAI Blog: workflow infrastructure, AI agents, and the future of work

StewAI BlogField notes from StewAI on AI workflow automation, agent reliability, deterministic recipes, and how modern teams turn ad-hoc work into versioned, executable infrastructure.

An Enum Has Four Answers. A Paragraph Has Infinite.: An honest case for structured output between LLM steps. The token-efficiency argument is debunked: compact JSON costs more than equivalent prose and providers bill the schema as input tokens. The real mechanism is answer-space collapse, measured across 114 StewAI recipes containing 267 enum fields with a mean of 3.6 permitted values. The reasoning-degradation debate is presented from both sides, including the finding that the most-cited anti-structure paper measured format restriction improving classification by roughly 19 points. Also covers three honest costs: schema conformance is not semantic correctness, field names act as instructions, and constrained decoding opens a jailbreak surface.

The Recipe Is the Ontology: A response to Frank Coyle's neurosymbolic talk. A StewAI recipe carries all three parts of an ontology: named versioned schemas as the concepts, a typed graph as the relations, and deterministic compute steps and gates as the axioms, with a schema registry making the vocabulary shared and importable by URN. The architecture matches Coyle's validator pattern exactly, but the symbolic layer holds a procedural model rather than a declarative domain ontology, which means it executes and does not infer.

We Paid 28x More for the Same Answer (And the Only Failure Was the Reasoning Model): A pilot benchmark on StewAI production: 15 runs, 4 models, 3 providers, 90 scored item-judgments on a source-grounded legal citation check with entirely fictional authorities. Three models across a 30x price range scored 100% on every unambiguous item; the only failing arm was a mid-priced reasoning model. The deterministic gate reached the correct routing decision in all 15 runs, including on the failing arm, because blocking aggregates over several items and cannot be flipped by one misclassification.

Thinking Work Is Made of Fourteen Blocks (Whether You Like It or Not): A catalog of fourteen reusable reasoning primitives extracted from production decision workflows: structured intake, isolated panel, checklist sweep, criterion scorer, mechanical aggregator, verdict rollup, consider-the-opposite, brief renderer, steelman balance, kill criteria, load-bearing assumptions, citation check, dedup cluster, and reference class. Each encodes a specific decision-science result, and the debiasing lives in the wiring rather than the prompt.

How AI Agents Work, What a SKILL.md Is, and Where StewAI Draws the Line: How the agentic loop works (reason, act, observe, with the model choosing the path at run time), what a SKILL.md is and how progressive disclosure loads a skill in layers, where agents and skills genuinely shine (open-ended, one-off, exploratory work), and a fair critique for high-volume known work: non-reproducibility, no pre-execution inspection, error compounding, and no statistical process control. Agents discover the process; recipes preserve it.

The Last Lag: What Three Centuries of General-Purpose Technologies Predict About AI's Payoff: Steam took 160 years from invention to measured productivity payoff, electricity 90, computing 55, the internet 30. The compression is measured, and a naive extrapolation puts generative AI's payoff around 2027. The one lag that never closes on its own is organizational redesign: unit drive for knowledge work.

They Approved 1.2 Million. The Outside View Said 3.3, and Refused to Be Sure.: Reference class forecasting as a deterministic workflow: the class is fixed before any evidence arrives, every comparable is validated in isolation against its source, the anchors are pure arithmetic, and a counting gate refuses to write a confident recommendation when the class is thin. M004 Outside View.

The Assembly Line for Thought: Three waves of industrial automation teach one lesson: the unit of progress was never the tool, it was the process. Chat is a better wrench, agentic AI a better craftsman; deterministic recipes are the assembly line, and statistical process control is how you trust them at scale.

It Got Four Experts to Agree, Then Refused to Call It Consensus: The RAND Delphi method as a deterministic workflow: anonymized submissions, a controlled feedback digest, isolated revisions, and convergence measured as a spread computed in code. On a real run, M003 Delphi Consensus narrowed a 15-month expert disagreement to 8 and honestly reported dissent instead of forcing agreement.

The Premortem: Assume It Already Failed, Then Explain Why: Gary Klein's premortem as a deterministic workflow: declare the plan already failed, generate failure causes through five independent lenses, ground and score each one, and emit a risk register with preventive mitigations. On a real launch plan, M002 Project Premortem surfaced thirteen distinct failure causes.

The Devil's Advocate Problem: Why Assigned Dissent Backfires: Four decades of research show committees converge too early and that appointing a devil's advocate often backfires, because assigned dissent gets discounted. M001 Committee Dissent encodes authentic, evidence-grounded dissent as a workflow: extract the assumptions, gather evidence for and against each in isolation, and assemble a decision brief nobody has to personally own.

The Model Was Never the Hard Part: Five Convictions Behind StewAI: Five evidence-backed convictions for why StewAI exists: atomic model judgment already works, enterprise value depends on context and workflow design, cost-first autonomy degrades quality, headcount is the wrong baseline, and the winning loop is fast iteration around structured recipes.

Do You Know How the Work Is Done? A 2x2 for Recipes vs Agents: A 2x2 for deciding between an autonomous agent and a deterministic recipe, based on whether you can specify the work and how often you will run it, plus the evidence that workflow redesign, not model choice, drives AI value.

The New Solow Paradox: AI Is Everywhere Except in the Productivity Statistics: Frontier AI matches experts on nearly half of professional tasks, yet forecast productivity gains are tiny. The 25x gap lives in the software layer between the model and the work, and value arrives only after organizations reorganize around the technology.

Humans Draw the Lines: What 106 Experiments Reveal About Human-AI Teams: The largest meta-analysis on human-AI collaboration found the average team performs worse than the best of either alone. The teams that win decide who does what before the work starts; the answer is structure, not better prompting.

The Silent Sabotage: Why 80% of Workers Run a 1944 Playbook: Maps the 1944 OSS Simple Sabotage Field Manual onto modern disengaged workplaces and shows how deterministic, auditable recipes make organizational sabotage structurally impossible.