VOL.2026.09.24 Today's call 3 min read
Thursday, September 24, 2026
agents are moving from talking to running
The clearest positive direction today: the value center of agents is shifting from conversation quality to orchestrated, reusable, schedulable process assets. AgentRun uses a DSL to turn agent calls into workflow definitions; ai-employees turns roughly 60 routines across 8 business roles into schedulable scripts. Both point at the same thing — what users want is not one answer but a process file they own and can run repeatedly.
For developers wiring existing agents into production. It offers a DSL that freezes ad-hoc call steps into repeatable workflow definitions, producing a reusable workflow config. Judgement: it targets the orchestration layer, not model capability; value depends on whether developers will maintain a separate process asset. Pricing, team and user scale undisclosed.
For small teams or individuals handling repetitive chores, it runs about 60 scheduled routines across 8 business roles in the browser, on Claude Code and ten other harnesses; the user ends up owning the process files. Judgement: open source plus multi-harness compatibility lowers switching cost, but it also means the moat is not in execution but in how well the templates fit a vertical.
For developers and teams needing Chinese and English speech turned into text in real time, the model takes a continuous audio stream and emits text on roughly an 80-millisecond clock, giving a streaming result that can feed subtitles, notes or downstream systems. Judgement: latency itself is becoming a sellable capability, with entry points in latency-sensitive legacy flows like support QA and meeting notes.
For brand, PR and content teams, it takes the brand or question set to monitor, inspects which sources AI answers cite, and outputs a traceable citation record. Judgement: as users ask AI directly instead of clicking search results, brand visibility is measured in citations rather than rankings — a newly emerging measurement need. Monitoring scope, update frequency and delivery format still need verification.
Radical Numerics feeds sequences and multimodal experiment data to a model that reasons step by step, outputting an understanding of new genomes, shifting the basis of judgement from literature search to reasoning over the sequence itself. Basecamp Research's public material only identifies it as a UK life sciences company using AI to speed up drug development and discloses the round's amount and investors; what data it ingests, who uses it at which node and what it delivers are not provided and remain unverified. CC is a family-facing AI agent extended to group use so several family members can take part; the materials it takes in and the actions it performs are undisclosed. ChatGPT added a voice-driven agent entry point on mobile for Pro and Plus subscribers, entering the queue for observation.
three things changing delivery conditions
The inference cost structure has changed. Reports indicate a 213:1 input-to-output token ratio in agent scenarios, with long context and multi-step tool calls as the main cost drivers. Agent products cannot copy the cost model of conversational calls; per-call pricing needs redoing.
The on-device surface is widening. On 23 September 2026 Qualcomm announced at its Snapdragon Summit that the Snapdragon X2 series expands to the Linux ecosystem, saying agentic AI PCs are accelerating. For AI applications, the hardware base for on-device agents is broadening. At the same time, in August 2026 Google set new memory-use limits for Android apps against a backdrop of AI data-center-driven hardware shortages; tightening platform memory constraints squeeze the space available to on-device AI features. Both directions exist at once, so on-device deployment needs tiering by device.
Healthcare AI compliance is taking shape. A UK commission published 44 recommendations on healthcare AI regulation on 22 September 2026, covering compliance and market-access frameworks. For healthcare AI applications this means clearer regulatory requirements before delivery, likely raising compliance cost and time-to-market; conversely, products that pass review gain an access barrier.
Real demand is shifting from what a model can do to whether a process can be frozen, who owns it, and what one run costs. Judgement: orchestration layers and process templates are the most concentrated opportunity band right now, but pricing and delivery costs are broadly undisclosed, so no revenue model can be concluded yet.