LoginSign Up
← Back to all sparks
Daily Brief · August 7, 2026

Context stops being a prompt: MotherDuck, Rootly and Customer.io ship it as a stored object

context-as-assetsemantic-layeragent-skillspolicy-enforcementstricter-defaultsmigration-tooling
Y
Generated 1h agoDrawn from 33 products

The lead

MotherDuck shipped Guides: an organization loads its metric definitions and join rules once, and every agent session writes SQL against the same semantics. The framing matters more than the feature: this is offered instead of better prompting. A dozen products made the same substitution today — what ships is not a smarter assistant or another tool, but a stored, permissioned description of how one customer's business works, which any model can read.

That is a narrower claim than it sounds, and a more defensible one. The model is rented, the tool surface is being trimmed, and an MCP server is unremarkable. What is left to own is the context. Tability states the division of labour outright: its Skills Library targets Claude and ChatGPT, and the company appears content for the reasoning to happen there while Tability holds the structured record. The record, not the reasoning, is the asset.

What moved

  • Context became an object, not a prompt. Rootly added CLI catalog sync from GitHub and Backstage so its AI gathers evidence itself across observability, code, infrastructure, flags and docs — read-only, nothing stored, no human pasting context into a chat panel. Customer.io moved prompt context out of individual actions into a workspace-scoped knowledge library. Chord grounds answers in the customer's own business definitions before Ask Chord may build an audience from them. HiBob retired its AI CV summary for a match score computed against criteria the recruiter defines.
  • Capability packaged as a named, assignable skill. NocoBase assigns skills per AI employee and builds them from its existing plugins, so the extension ecosystem — not a separate tool registry — defines what agents can do. Docling added agent skills alongside its new video pipeline. ElevenLabs goes furthest: ElevenAgents has branches, rebase, merge previews, drafts and per-test tool mocks, and Procedures compile into workflows. Honeybadger removes the other half of the expertise: natural-language search writes the BadgerQL, and learned baselines replace threshold tuning.
  • Enforcement moved to the interception point. Speakeasy checks prompt policies at the LiteLLM proxy: a violating prompt dies before inference rather than being reported after. AWS Machine Learning gave AgentCore temporal policies that evaluate an agent's session history rather than a single call, plus gateway rate limits scoped by JWT claim, and open-sourced the policy language behind them. Zed sandboxed its agent's terminal commands and web fetches. Prowler let Lighthouse manage users and roles over MCP, with RBAC the only guardrail — behind the subscription.
  • The upgrade changes your configuration. gRPC 1.83 turns on post-quantum key exchange in TLS by default, no flag; SoftEther VPN added the same X25519MLKEM768 exchange. Protocol Buffers 36.0-rc1 flips default symbol visibility to strict for Edition 2026, Apache ActiveMQ keeps walking its defaults inward (servlet off, XBeanBrokerFactory blocked, AMQP frames capped at 10 MB), and Dovecot 2.4.4 permanently drops root in three services. Where that breaks things, the migration tool ships beside it: Apache Maven is one candidate from GA, with mvnup to rewrite pom.xml files and a 3.10 line that back-ports Maven 4's resolver semantics. OctoPrint 2.0 demotes serial to one connector among several and warns plugin authors they will break, as does Craft CMS 6.0 alpha 14 on its way onto Laravel's container. Gradle 9.7 RC3, Terragrunt v1.1 and Tekton's new v1.15 LTS are stabilisation, not features.

Sectors today

Devtools (39 products) carried both halves of the day; development (22) was almost entirely release engineering. Collaboration (19) split between SiYuan's 3.8 branch adding OIDC, semantic search and MCP, and Typst widening past PDF into HTML and MathML. Project management (12) is where agents took operational steps under human approval — Tability, Atlassian's Summer Jira release, OpenProject 17.7 annexing staffing. Analytics (11) was MotherDuck and Chord; design (7) belonged to shadcn/ui becoming a layer over several component bases. Video conferencing (13) had one real move, Wowza hosting NVIDIA's synthetic-video detector. HR-recruiting (10), ecommerce (9), finance (9), marketing (7) and customer-support (7) were thin on releases and thick on blog feeds.

Watch tomorrow

If context is the asset, the next question is whether it stays proprietary. AWS open-sourcing its agent policy language and Tability pointing its Skills Library at other vendors' assistants cut against lock-in; MotherDuck's Guides and Chord's business definitions do not. Watch whether these stored objects acquire the review machinery ElevenLabs gave its agents: a shared semantic layer anyone can edit is a correctness problem before it is a governance one. Two dates are fixed: Buildkite Agent v4 goes stable on 1 September, and GitHub has closed Spark to new users. On crawl quality, roughly 40 of today's 188 feeds carried no releases: Metricool, Celoxis, Recruiterflow and Intermedia are content programs, and Firefly III still publishes nightly build tags. Customer.io is dual-rowed again; one row is cited.