LoginSign Up
← Back to all sparks
Daily Brief · August 26, 2026

The human moved back inside the run: Subbly, Tana and Copilot ship mid-flight controls

mid-run-steeringagent-interruptiondefault-inversiondeprecation-datescontinuous-approvalmcp
Y
Generated 7d agoDrawn from 49 products

The lead

Six days ago the pattern here was the human leaving the trigger — agents that start themselves on a schedule or an event. This window runs the other way. Subbly made its AI Builder interruptible: a prompt queued mid-run is inserted between steps rather than waiting for the run to end. Tana turned AI proposals into something you edit in place against the current values, rather than accept or reject whole. Workato gave AIRO a stop button. And GitHub Copilot put agent sessions in Slack and Microsoft Teams, turning a session that bound one developer to one editor into something a channel can watch and redirect.

Taking the person out of the start is what made the middle the only place judgement can land. The change is sharpest at Tana: accept-or-reject treats a model's output as final or void, while editing in place treats it as a draft — a different object, with a different review cost. Subbly names the price, in credits spent on a run going the wrong way. The same shape appears with no AI in it at all: NATS 2.15's preview can cancel an in-flight stream move, evacuate a peer, or rescue a metalayer that has lost quorum — operations that previously required peer-remove and hoping. A long-running operation stopped being atomic.

What moved

  • The default became the decision. ApexCharts 7.0 leads with a removal: nine features leave the default bundle for explicit imports, taking licence-gated code in the default path from 24% to zero. Tree-shaking already worked — it just required reading the docs. DatoCMS drops freeform admin theming for new projects in favour of one hue at WCAG AA contrast. OpenMetadata 2.0 switches its MCP server on by default, and moves embedding configuration so semantic search stops working silently unless the new block is set.
  • The end date shipped with the release. Apprise 1.13.0 announces the v1 line ends there and tells downstream projects to pin below 2.0.0. Dosu changes its open-source terms on 1 September, saying roughly 96% of projects see nothing. Infisical began flagging native integrations for deprecation in-product and by email. Rocket.Chat deprecates LDAP and SAML on non-Premium workspaces and names 9.0.0 as the enforcement point.
  • A standing approval that expires. Speakeasy re-checks approved MCP servers daily for permission changes, on the reasoning that a one-time verdict goes stale without saying so. Hatz AI moved tool credentials to the admin who resells the platform, leaving personal connections dormant and badged Superseded. Mailtrap caps what a stolen credential can reach with a recipient-domain allowlist.

Sectors today

  • Devtools (20): incident.io and Speakeasy build the surround for agents; Biome grinds.
  • Development (16): QuestDB tiers storage to object stores; NATS previews reversible cluster operations.
  • Analytics (11): ApexCharts leads; Neo4j consolidates tenants, Polars publishes its 2.0 migration guide.
  • Communication and messaging (11): Chatwoot absorbs WhatsApp templates; Grain labels speakers by voice.
  • AI assistants (9): Claude ships memory with its controls; Baseten sells capacity timing.
  • Collaboration (8): Tana leads; GitHub thickens governance while Hive widens surfaces.
  • Marketing automation (8): Customer.io, Sender and Gumloop moved reporting or triggers outward.
  • Customer support (6): DocsBot AI picks up the phone; Respond.io auto-answers missed calls.
  • Project management (6): ClickUp wires MCP into Super Agents; Camunda pauses features for defects.
  • HR and recruiting (6): Fountain's Cue reaches outside tools; four of six feeds are pure content.
  • Design (5): Storybook makes component metadata agent-readable; Oxygen is six betas into MCP repair.
  • Ecommerce (5): CartFlows hands order-bump choice to a model; Sendcloud widens carriers.
  • Marketing (5): Privy picks its own A/B winner; Meltwater ties AI visibility to revenue.
  • LMS and edtech (3): Code.org's Python Lab grows audio and images; TAO Testing adds audit logs.
  • Video conferencing (3): PeerTube federates moderation; Asterisk remains a backport ledger.
  • CRM (2): folk opens its interaction history; Phorest strips steps from the salon desk.

Watch tomorrow

InvokeAI 6.14.0 cut a final after two candidates dating to 31 July, closing one long train and leaving others open: Gradle has cut two 9.8 milestones while 9.7 still has no final, OctoPrint sits at 2.0 rc5 on unexplained error states, and PeerTube's 8.3.0 candidate is the largest release in its window. The dated commitments are Dosu's 1 September terms change and Rocket.Chat's 9.0.0, which now carries the LDAP and SAML licence requirement. On the day's main thread, the question is whether mid-run correction gets recorded anywhere: Workato says its thumbs up/down feedback will feed batch runs and automated grading, which would turn a one-off save into training data. Several feeds again carried no product signal — Jobvite, iCIMS, Comm100, Metricool and WebinarJam are content operations, not changelogs, and Rancher publishes tags whose notes itemize nothing.