THE DALTARIS RADAR · November 13, 2026

The week review became the bottleneck.

Coding agents added 44% more code at 655 firms and shipped at most 31% more software, because approval now takes 56% longer. Anthropic tells the SEC its models may soon speed up AI research itself, the best agent finishes 12% of real professional tasks, NVIDIA open-sources the semantic layer, and a 2.8-trillion-parameter model goes open-weight. Also this week: Confluent builds the Streamhouse, Postgres 19 slips because AI keeps finding bugs, and OpenAI's Evals platform goes dark.

#0

Featured: the reviewer is the bottleneck

What happened: Two Harvard PhD candidates, Fiona Chen and James Stratton, posted the current version of 'Artificial Intelligence in the Firm: Bottlenecks in Software Production' on November 1, built on Jellyfish data from 655 firms, 720,542 workers and 445 million work events between January 2021 and August 2026. After firms adopted coding agents, lines of code rose 44.0%, commits 36.0% and pull requests 35.7%, but shipped software rose at most 30.7%. Time from pull request submission to merge rose 56.1%, from a 3.56-day baseline: code waited 59.9% longer in the backlog and spent 58.5% longer in active review. The same week, GitHub's Octoverse reported that over 40% of pull requests now start as an agent's first draft, about 110 million in September, up from 17% in March.

Why it matters: The flagship's mechanism in one line: generation is infinite, verification is the constraint. Review is done by the person who already holds the system in their head, 82.8% of reviews in the data. AI review tools did not clear the queue: 79.1% of firms had adopted them. Read the numbers with care: the paper is a working paper, not yet peer-reviewed, and GitHub's figures are vendor research. The direction of both is the same. Visual: reviewer-bottleneck-2026-11-13-v1.png, scheduled as the Friday X post (draft 11755522).

What to do: Split time-to-approval into its two halves and put both on the wall: backlog wait and active review time. If backlog is the problem, you lack reviewers. If active review is, you lack tests.

#1

Anthropic tells the SEC the loop is closing

What happened: Anthropic filed Amendment No. 1 to its S-1 on November 2, adding a risk disclosure: based on internal estimates for future models, the company may imminently meet its Responsible Scaling Policy definition of 'dramatically accelerated progress attributable to AI', roughly doubling the pace of AI progress or replacing its research workforce. No current model has been designated as crossing the threshold. The same filing says Claude wrote over 90% of contributions to Anthropic's codebase in September 2026, and average daily code per engineer in the third quarter ran over 45 times the 2024 average. Anthropic also says it recently delayed pre-training a frontier model by more than a month to harden security.

Why it matters: The company building the model is telling its own investors the loop may be closing, and it is the same loop as this week's flagship: code volume is compounding while the human checkpoint does not scale. The 45x figure measures code volume, not productivity, and Anthropic says so itself. Read the disclosure as a governance document, not a capability announcement.

What to do: If your team ships AI-written code, ask who reviews it and what the queue looks like. That is the same question as the flagship, asked of your own repo.

#2

The best agent finishes 12% of the work

What happened: SemiAnalysis published GDPx on November 2: 206 long agent tasks across 11 professional fields, chip design to law, with up to 8 hours per task and binary pass/fail grading wherever possible. The best model, Claude Opus 5.5, finished 12.0% at $51.41 per task. Gemini 4 Carbon scored 10.9% at $16.29, GPT-6 Astra 9.6% at $53.99. Every model scored zero on finance and legal tasks under strict grading. SemiAnalysis says safety classifiers pulled down scores for Opus 5.5 and GPT-6 Astra in cyber, biology and alignment.

Why it matters: This is the third costume for the same finding as Kaggle and AgentSWE: the demo is not the delivery. The useful number is not 12%, it is $428 per successful outcome for the leading model. Budget the task, not the benchmark.

What to do: Before you put an agent on a professional task, run it on ten real cases and count completions, not demos.

#3

The semantic layer goes open source

What happened: NVIDIA released Auto Ontology under Apache 2.0: a framework that builds a semantic, graph-based layer from enterprise metadata and serves it to agents through MCP, a CLI and an API, so agents query certified business definitions instead of raw schema. It works with Databricks, Postgres, Snowflake and DuckDB, stores the catalog in Postgres with pgvector, and picked up roughly 8,900 GitHub stars in its first days.

Why it matters: Retrieval finds the table called revenue. A semantic layer tells the agent what revenue means, and which definition Finance will defend. This is the same argument as Snowflake Horizon Context last week, from the open-source side: the agent reads your definitions next, so their quality is infrastructure.

What to do: Write down your three most disputed metric definitions. If they live in three places, your agent already has three answers.

#4

A 2.8T open-weight model, fourth on the index

What happened: Moonshot released the weights for Kimi K3.1 Pro, a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active per token, and a 1-million-token context window. It scores 55 on the Artificial Analysis Intelligence Index, fourth overall. Moonshot stripped cybersecurity-related data from training before the open release: on its in-house vulnerability benchmark the released model scored 81.0%, against 90.5% for a checkpoint without the removal.

Why it matters: The open-weight frontier keeps closing on the closed one, and the release itself is now a governance decision: what the model cannot do is chosen before the weights ship. Every headline number traces to the lab or a single index; no independent harness has rerun the released weights yet.

What to do: If you self-host, price the serving bill before the license. A 2.8T model is free to download and expensive to run.

#5

Signal: Confluent builds the Streamhouse

What happened: Confluent announced its Streamhouse architecture at Current 2026 on November 4: a data architecture that keeps business context current for agents, built on Kafka, Flink, governance and a new Query pillar. Lightning Tables, now generally available on AWS, serves the current state of the business as a read-only layer. In Confluent internal testing, data moves from creation to serving in about 100 milliseconds, with query latency of 10 to 100 milliseconds and throughput above 1,000 queries per second.

Why it matters: The lakehouse tells the agent what happened. The Streamhouse tells it what is true right now. Same argument as Auto Ontology and Horizon Context, from the streaming side: stale context is a data problem, not a model problem.

What to do: Measure the age of the data your agent reads. If it is a nightly batch, it is answering yesterday's question.

#6

Signal: Postgres 19 slips because AI keeps finding bugs

What happened: PostgreSQL committer Robert Haas wrote on October 30 that version 19 will slip to November 19, outside the project's usual mid-September to mid-October window, because people using LLMs are finding real bugs in a 30-year-old codebase faster than committers can fix them. Backports to the release branch rose from 724 to 1,169 in the same window year over year, a 61.5% rise. Three major features were pulled late in beta, between August 27 and September 15.

Why it matters: The flagship's mechanism, running on the world's most careful database project: finding got cheap, verifying did not. The same queue, the same single human at the end of it.

What to do: Postgres 19 RC2 ships November 12, the same day version 14 reaches end of life. If you still run 14, the fixes stop there.

#7

Signal: OpenAI's Evals platform goes dark

What happened: OpenAI's Evals platform went read-only on October 31, and its dashboard and API shut down on November 30, along with Agent Builder and reusable prompt objects. OpenAI points teams to Promptfoo, an open-source eval framework, with a migration guide for recreating dataset evals. Multi-turn tests have to be rewritten by hand.

Why it matters: While everyone talks about evals, the evals platform itself is being retired. The deadline that matters is the first one: read-only means your suite keeps returning a confident score while it stops describing your product.

What to do: Export your eval definitions and results before November 30. A suite you cannot add a case to is a snapshot, not a gate.

SOURCES

✦ The Daltaris Radar lands weekly. Follow @Daltaris on X to get each issue first.