Before revealing AlphaCRO’s results, a centered question asks how a harness built for Coframe handles coding tasks it wasn’t optimized for, using the same agent and harness without changes. The phrase “it wasn’t optimized for?” stands on its own line in warm coral, making that premise clear from a distance.
The ranking first introduces “AlphaCRO” at 80.0%, without revealing its mode. The question and mode cues type in, and the existing result gains a typed “Prototype” suffix. Only Prototype’s faster-and-cheaper label appears at that stage; Build’s label arrives with its 81.1% result. The final cue reads “One harness.” and “Two modes.” on separate lines. The following charts compare their accuracy, cost and time.
The finale first shows AlphaCRO powering Mark-17 today. Next types “The same harness. Soon, for every team.” A second Next reveals the company-wide uses, including CroBot, under “Coming soon · across Coframe.”
The infrastructure view describes the Agents SDK’s role as “Shared sessions. Live responses.”
Taller, wider cost and time plots make the results easier to compare at a distance, with one Accuracy axis title, more space between labels and the same scales throughout the reveal.
After Build appears on the cost and time plots, Next goes straight to the finale. The extra bottom summary and its click are removed.
The completed robot lifts both fists beside its head in a two-arm flex, with neither hand hidden behind its face.
The Mark-17 announcement’s future messages stay hidden from screen readers until the presenter reveals them; printed slides include the complete announcement.
After Mark-17 lands, the announcement reads “Please welcome…” before the next slide introduces its name.
State is illustrated with a concrete work record—file edited, tests failed, fixing the error—instead of an ambiguous backpack. The next step shows the same record as a saved checkpoint to explain durability.
The presentation opens with “Agent Eng Presents”, then the two video slides announce Mark-1’s retirement and Mark-17’s arrival. “Trust me.” crossed out by one light rising stroke above “Show me.” follows the announcement. Both opening cards omit slide numbers. Next introduces “Harness.”, then two clicks ask what it is and what it takes to build. “Let’s build one.” begins the model-building story. The deck has 18 slides. Prototype’s results show 80.0% accuracy and 16.0 minutes average time per task.
alphacro-harness builds a harness around a small robot, one layer at a time, then zooms out to its Cloudflare and Kernel infrastructure. Eighteen visual Show and Tell slides include the retirement announcement, engineering work and benchmarking. An empty stage introduces Mark-17, then the heading erases and types an invitation to welcome AlphaCRO, our internal codename, with “The harness behind it” typed beneath. Prototype appears before Build against familiar harnesses. Large rankings and side-by-side accuracy/cost/time charts keep the same scales as AlphaCRO enters, with fewer numbers and stronger visual emphasis on its results. The ranking scale and caption appear with the results. The finale previews company-wide availability. Next controls each reveal; the retirement scene keeps its original idle loop, while arrival and hero clips play once. Includes speaker notes and a printable deck. Returning from print preserves the active slide without replaying its entrance.
