Week of 31 August 2026 — DuckDB joins AWS, and the case for context engineering

Compiled on 22 September 2026 for The Brief's launch archive, from sources published between 31 August and 6 September 2026.

Four stories that mattered

AWS completes its purchase of the team behind DuckDB

Amazon's deal for DuckLabs, the roughly 30-person Amsterdam company founded by DuckDB creators Hannes Mühleisen and Mark Raasveldt, closed on 31 August, and the team formally joined AWS on 1 September. DuckDB itself — the fast-growing embedded analytics database favored by developers who want to query large datasets without standing up a warehouse — is not part of the sale: its intellectual property stays with the independent, nonprofit DuckDB Foundation, spun out of the Dutch research institute CWI, and it remains MIT-licensed. The founders said DuckLabs, self-funded until now, could not keep building sales and support infrastructure around more than a million daily downloads without it becoming a distraction from engineering. As part of the deal, the Foundation is adding a technical advisory board and opening the extension ecosystem to third-party developer signing.

Analytics teams that have quietly embedded DuckDB into pipelines over the past few years now have a large single vendor sitting next to their dependency, even though the license and governance structure has not changed yet.

Source: Help Net Security, DuckDB stays open source while the team behind it goes to work for Amazon

A BARC study puts a number on "context engineering" for AI agents

A global survey of 285 data, AI, IT and business stakeholders, run by the Business Application Research Center and sponsored by DataHub, found that organizations it classified as "context leaders" — the 42% that had implemented, formalized or optimized six foundational practices such as data integration, retrieval, federated metadata and semantic modeling — were four times more likely to also qualify as "AI leaders" (49%, versus 12% of everyone else). Data quality and preparation was the most commonly cited obstacle, named by 49% of respondents, ahead of model limitations (29%) and governance gaps (25%).

The study is vendor-sponsored and self-reported, but the core claim — that agentic AI succeeds or fails on the unglamorous work of organizing and connecting data, not on model choice — matches what most working analytics teams already suspect.

Source: DataHub, New BARC Study: Context Engineering Leaders Lead in AI

US employers added 162,000 jobs in August, well above the recent trend

The Bureau of Labor Statistics reported that nonfarm payrolls rose by 162,000 in August, against a prior 12-month average of roughly 31,000, while the unemployment rate held at 4.1%. June and July figures were revised up by a combined 55,000 jobs. Food services and drinking places added 59,000 jobs and local government education added 42,000, while information-sector employment fell by 23,000. Average hourly earnings rose 0.3% for the month and 3.1% over the year, to $37.75.

For anyone building economic or workforce dashboards, the two revisions are the detail worth flagging to stakeholders: a single month's headline payroll number is provisional for at least two more reporting cycles.

Source: Bureau of Labor Statistics, The Employment Situation — August 2026

The world's population of over-65s has overtaken its under-5s, Census Bureau says

The Census Bureau's "An Aging World: 2025" report, prepared with the National Institute on Aging, found that the global share of people aged 65 and older passed the share aged under five for the first time in recorded history, and projects that share will nearly double, from 10.5% to 19.6% of world population, between 2025 and 2060. The United States currently ranks as the 48th-oldest country (18.9% of its population is 65+) but is projected to fall to 110th by 2060 as other countries age faster; South Korea is projected to become the world's oldest country, at 41% elderly.

This is background context, not a KPI, but it belongs on the radar of any team modeling long-run consumer, healthcare, workforce or pension data.

Source: U.S. Census Bureau, Global Older Adult Population Projected to Nearly Double by 2060

Tool moves

• Domo's $400 million sale of substantially all its assets to Progress Software cleared US antitrust review (early termination of the Hart-Scott-Rodino waiting period) on 25 August; Domo said it still expects the deal to close by the end of September. Source: StockTitan, Domo Announces Timing for Second Quarter Fiscal 2027 Financial Results and Provides Transaction Timing Update
• Databricks's Genie One can now search the public web for current information and, as of 1 September, can run on OpenAI models and Anthropic's Claude Fable 5.1 hosted directly on Databricks, alongside its existing model options. Source: Databricks, release notes — September 2026
• Databricks added the ability to share metric views with users in other metastores (OpenSharing) from 3 September, letting governed metric definitions cross organizational boundaries without duplicating them. Source: Databricks, release notes — September 2026

One how-to

A starter checklist for "context engineering." This week's BARC study argues that what separates organizations getting real value from AI agents is not the model but the metadata, structure and plumbing around it — what the study calls context engineering. A practical way to start:

1. Inventory your semantic layer before your model. If "active customer" has three SQL definitions across dashboards, an agent will confidently pick one and be wrong.
2. Separate retrieval from generation. A retrieval-augmented setup pulling from a governed vector database or knowledge graph is auditable; an agent that free-associates from raw text is not.
3. Design for a shrinking context window, not a growing one. Dumping whole documents at a model instead of retrieving the relevant slice is slower and, per the study's top challenge, more likely to surface stale data.
4. Put a human checkpoint in the orchestration layer until the agent's error rate on your own data is measured, not assumed.

None of this needs a new platform. It needs treating context as an engineering discipline, not a prompt-writing exercise.

One number

$37.75. That was the average hourly earnings for US private-sector workers in August 2026, per the Bureau of Labor Statistics — up 0.3% on the month and 3.1% over the year. It is a broad, timely signal of wage pressure across the economy, watched closely by anyone building labor-cost or pricing models. What it does not tell you: it is an average across very different jobs and industries, it is not adjusted for inflation (so a 3.1% nominal rise is a smaller real-wage gain once price increases are subtracted), and it says nothing about hours worked, which also affects take-home pay.

Source: Bureau of Labor Statistics, The Employment Situation — August 2026