AI agents break software out of its app shell
Summary
The recurring shift is from software as a destination to software as a set of capabilities agents can call: data, reliable execution, and accumulated workflow knowledge become more important than owning the screen. That raises the premium on interoperability and proof—buyers need visibility into real usage before consolidating vendors, while builders need evidence, risk controls, and reviewable processes rather than simply adding an agent layer. Advertising is beginning to follow the same interface transition, adapting commercial measurement and presentation to AI-mediated use.
Show HN: AI search for every photo and every frame of video on macOS
SCM is a local-first macOS app that indexes photos and video frames without accounts, cloud uploads, or remote inference. Its visual search combines embedding similarity with filename and phrase boosts, per-model score calibration, duplicate filtering, and explanations of why a result matched; it also handles CJK queries with overlapping bigrams. For video, it detects scenes with ffmpeg, indexes midpoint frames, limits each video to three matching scenes, and opens hits at the relevant timecode. OCR and Whisper provide separate literal-search fallbacks that remain usable while the AI engine is offline, while an opt-in local LLM can answer questions from extracted evidence with clickable citations.
Read the source →The end of the app era: What comes next?
The video argues that agents make a software application's interface less defensible: a worker's agent can read contracts, sales notes, emails, and records, then perform the work that formerly required opening several tools. It separates the stack into data, workflows, application interface, and agentic layer, contending that vendors can lose the interface while retaining value through dependable data and execution—much as DoorDash can take orders through another assistant via MCP. Buyers should learn how teams actually use agents before changing vendors, because individual workflows and accumulated agent context create real switching costs that central procurement may not see. Sellers need more than an MCP endpoint or an “AI” claim: they need to show why their data, workflows, and specialized agents improve business outcomes while preserving model choice. The advice to employees is to make effective AI workflows visible to managers rather than treating them as private productivity tricks that can be accidentally broken by a vendor switch.
Read the source →New Skills! v1.3 brings /pr, /implement-spec, and /retro
Matt Pocock's Skills v1.3 adds an implement-spec workflow that turns a large specification into tickets, executes them in bounded agent sessions, merges through an integration branch, runs review, and produces one reviewable PR. The aim is to avoid long single-agent runs that hit context limits or drift, while leaving room for teams to adapt the orchestration to stacked PRs or other patterns. Its PR skill uses a structured body with a visual summary, before-and-after evidence, and an explicit assessment of reversibility and blast radius so reviewers can judge risk quickly. The new retro skill examines completed agent sessions for missed guardrails, navigation problems, tool/token waste, context loss, and unsafe actions, then suggests improvements for a human to choose from. Pocock explicitly recommends sampling sessions and keeping retro human-in-the-loop rather than automating every proposed repository change.
Read the source →Executive Briefing: Agents Split Your Software Into 3 Parts. Which Should You Keep Paying For?
The article says that AI agents are already changing software strategy through the everyday decisions employees make about which data, workflows, and applications they trust or bypass. It frames a typical subscription as a bundle of data, interface, and workflow, then argues that agents let customers unbundle those pieces: they may discard screens while still needing records, approval rules, or execution. Leaders should surface existing agent use inside teams before standardizing tools or renewing contracts, and decide deliberately where owning a workflow creates advantage versus where buying a commodity service still makes sense. For software vendors, “we have AI” is not enough; they must explain how they will continue delivering value as models improve and customers demand more portability.
Read the source →Building advertising for the way people use AI
OpenAI announces a new visual advertising format for ChatGPT. It also says it is expanding advertiser measurement tools, attribution partnerships, and brand-suitability controls. The short announcement positions the changes as adapting advertising to how people use AI, but provides no further implementation or performance details in the supplied text.
Read the source →