TheFocus Labs No. 9 · Soft Clay
TheFocus.AI Press · Bound Monographs 3 editions in print

State.

Periodic, multi-chapter syntheses on the state of AI engineering — pairing field reporting from major industry summits with internal benchmarks from TheFocus.AI.

  1. State. № 03
    AI ENGINEER CODE SUMMIT · NOV 2026
    FIG. 03 // RETHINKING THE FACTORY SDLC NOV 2026 01 · SPECIFY Requirements & Specs • Demand discovery • Expected-pain spec • Human spec layer 02 · FACTORY Software Factory Code cost → $0 Cloud VMs + bash 03 · VERIFY Frozen Test Oracles • Read-only tests • AST & type gates • Video/trace proof BROWNFIELD TRAFFIC & TELEMETRY LOOP REBUILD SPEC (01) + VERIFY (03) AROUND FACTORY STATE № 03
    The San Francisco Report
    November 2026 7 CHAPTERS
    Latest Edition · № 03

    The San Francisco Report

    CONFERENCE TAKEAWAY

    Software factories force a complete rethink of the SDLC — changing how features get specified going in, and how frozen verification gates what comes out.

    AI Engineer Code Summit, November 2026 · rethinking requirements, feature specification, and frozen verification around the software factory

    Across 33 talks at the AI Engineer Code Summit in San Francisco (November 2026) and internal benchmarks at TheFocus.AI, teams running production software factories converged on a clear architectural lesson: once autonomous code generation drops toward zero marginal cost, the traditional SDLC breaks unless you rebuild both ends around the factory — upstream demand discovery and expected-pain feature specification going in, downstream read-only test oracles and deterministic AST/type gates coming out, L2/L3 organizational maturity under small full-stack pods, plus inline decision models, cloud VM sandboxes, and modeled data warehouses.

    1. From the Editor Rethinking the SDLC Around the Software Factory
    2. § I Decision Models Inside Application Code
    3. § II The Agentic SDLC and Organizational Maturity
    4. § III Specification and Verification in Existing Codebases
    5. § IV Personal Assistants Move to Cloud VMs
    6. § V Data Modeling, Pre-Indexed Search, and Memory Consolidation
    7. § ∞ What Remains Scarce
    Read STATE № 03 → Print edition 33-talk archive ↗
  2. State. № 02
    AI ENGINEER NEW YORK · OCT 2026
    FIG. 02 // CROSSING THE 95% DIFFUSION WALL OCT 2026 RAW AGENTS Unbounded Pilots • 95% stall rate • 80% wrong OK’d • Stale context • Prompt wishes • KV cache burn 50% CODE · 5% BIZ 01 · BESPOKE EVALS Answer-First & Process Dated web · Behavior specs 02 · RUNTIME GOVERNANCE Policy DAGs & Ontology Scoped tools · Golden data 03 · SANDBOX & ECONOMICS Linux VMs & KV Routing Open weights · Trace → Skill PRODUCTION Bounded Autonomy • Audited proof • Golden sources • Scoped vaults • 10–20× cheaper • Fleet managers GOVERNED SCALE TRUST INFRASTRUCTURE CROSSES THE 95% GAP STATE № 02
    The New York Report
    October 2026 6 CHAPTERS
    Edition № 02

    The New York Report

    CONFERENCE TAKEAWAY

    Half of large enterprises run coding agents, yet only five percent have autonomous agents working anywhere else in the business — crossing that 95% diffusion wall takes bespoke evals, runtime policy governance, and real tokenomics.

    AI Engineer New York, October 2026 · crossing the 95% enterprise diffusion wall with bespoke evals, runtime governance, and real tokenomics

    Thirty-one recorded sessions from AI Engineer New York in Midtown Manhattan (October 12–14, 2026), where frontier labs met hedge funds, private-credit desks, insurers, and infrastructure builders. While 90% of large enterprises have rolled out general LLMs and 50% run coding agents, only 5% have autonomous agents working in non-engineering operations—stalled by 80% human cognitive surrender on wrong outputs, noisy LLM judges, conflicting golden data sources, and runaway KV-cache bills. Crossing that gap took a three-layer architecture: answer-first domain evals and behavior specs, runtime policy DAGs over bitemporal ontologies, and VM-sandboxed fleets paired with prefix-preserving token economics.

    1. From the Editor Crossing the 95% Enterprise Diffusion Wall
    2. § I Bootstrapping Domain Evals and Verification Without Public Benchmarks
    3. § II Bounded Autonomy, Policy DAGs, and Governed Financial Workbenches
    4. § III Sandboxed Computers, Institutional Context, and the Shrinking Harness
    5. § IV Agent Managers, Self-Improving Skills, and Real Tokenomics
    6. § ∞ Beyond the Text Window — Continuous Control and World Simulators
    Read STATE № 02 → Print edition 31-talk archive ↗
  3. State. № 01
    AI ENGINEER MIAMI · APR 2026
    FIG. 01 // ENGINEERING WAS THE FILTER APR 2026 BEFORE · SCARCE ENGINEERING ACTED AS TASTE FILTER 100 IDEAS HARD TO BUILD FEW SHIPPED NOW · FILTER REMOVED — EVERYTHING GETS BUILT EVERY IDEA GOOD + BAD NO FILTER Unfiltered Flood 1st-draft code · Slop Taste is now scarce Identity collapses SHIPPING FRICTION USED TO DO THE FILTERING STATE № 01
    The Miami Report
    April 2026 4 CHAPTERS
    Edition № 01

    The Miami Report

    CONFERENCE TAKEAWAY

    The difficulty of software engineering used to act as a hidden filter against bad ideas — and we just removed it.

    Two days at AI Engineer Miami · engineering stopped acting as the filter, and the patterns haven't refrozen

    A dispatch from AI Engineer Miami, April 2026. Day One uncovered what implementation friction had been hiding: engineering difficulty used to act as an implicit taste filter keeping weak ideas from getting built (and propping up developer tribal identity), and removing that bottleneck unleashed an unfiltered flood of first-draft software as Gen-3 coding-agent SDKs reframed the unit of shipping. Day Two traced three populations — people, companies, and agents — scrambling for new patterns at once.

    1. From the Editor When the Filter Comes Off
    2. § I The Forces Have Moved
    3. § II Learning Agency: People, Companies, Agents
    4. § ∞ What Is an Agent?
    Read STATE № 01 → Print edition
Loading archive index…
↑ ↓ navigate ↵ open ⌘K toggle