4 AI drops worth watching: September 29
Vespper: ships an MCP for reliable Word document editing
Vespper, a YC F24 startup, launched a Docx MCP server on September 28 that lets AI agents edit .docx files without the corruption and formatting loss common to other approaches. The team, founders Dudu and Topaz, built the tool around a fine-tuned model and a round-tripping method that converts Word’s verbose OOXML structure into something an agent can edit safely, then converts it back. Vespper says the product is 3x faster and 2x cheaper than the closest alternative, and more accurate, though those figures come from the company’s own launch post rather than an independent benchmark. The founders previously built an AI document editor for pharma companies before starting Vespper.
Word documents remain the actual deliverable in legal, finance, and healthcare work, and agents that mangle tracked changes or numbering are a real blocker for vertical AI products in those spaces.
The take. The problem Vespper describes is specific and well-argued: agents burn context fighting Word’s XML internals instead of doing the actual legal or financial reasoning task. Round-tripping is a reasonable bet if it holds up on real templates with headers, footnotes, and nested styles, which is where lossy conversion tools like pandoc typically break. The 3x/2x claims are unverified marketing copy until a third party runs the same test suite; watch for legal-tech teams like Harvey’s engineering blog to reference it if the approach actually reduces their in-house maintenance burden.
Definitions:
- MCP (Model Context Protocol): a standard that lets AI agents call external tools, in this case Word document editing functions.
- OOXML (Office Open XML): the underlying zipped XML format that .docx files are built from.
GPT-6 Astra: Roboflow calls it the strongest vision model it has tested
Roboflow ran OpenAI’s GPT-6 Astra through its Vision Evals benchmark and published results on September 18. At low reasoning effort, Astra scored 82.1% mAP@50 on object detection, 5.4 points ahead of Qwen3.8 Max and 13.7 points ahead of GPT-5.6 Sol, with no competing model matching that score even at high reasoning effort. The model also handles box prompting (finding objects from example boxes rather than class names), cross-image prompting, and segmentation by returning polygon vertices instead of bounding boxes. Roboflow has made Astra the default model behind its Auto Annotate feature.
OpenAI built Astra’s release around computer-use tasks like locating and reading interface elements, and those same skills carried over into general-purpose vision benchmarking.
The take. This is a third-party benchmark, not a vendor claim, which gives the numbers more weight than a typical release post. The auto-annotation angle is the load-bearing detail: if a general vision-language model can now out-detect specialized detectors on common categories, the economics of building custom CV pipelines shift toward using Astra as a labeling layer and reserving fine-tuned models for the long tail of company-specific objects. Watch whether Roboflow’s Auto Annotate defaulting to Astra becomes the template other annotation tools copy over the next two quarters.
Definitions:
- mAP@50 (mean Average Precision at 50% IoU): a standard object-detection accuracy score; higher means fewer missed or misplaced boxes.
- reasoning effort: a setting that controls how much computation a model spends thinking before answering.
Imp: a full port of DSPy to Elixir’s BEAM
Imp is an open-source Elixir library, available on hex.pm, that reimplements DSPy’s declarative language-model programming model on the BEAM virtual machine. It carries over DSPy’s core pieces: typed signatures, prediction modules, optimizers including GEPA, MIPROv2, SIMBA, and fine-tuning/GRPO, plus agent loops built with Imp.react and retrieval support. The project has reached 230 GitHub stars and merged a 0.7.0 release as of September 29. Because agents in Imp run as supervised OTP processes rather than in-process function calls, they keep independent state, can be monitored, and restart under a supervisor like the rest of an Elixir application.
DSPy turns prompt engineering into a program you can measure and optimize instead of hand-tuning by trial and error; Imp brings that discipline to a runtime built for long-running, concurrent processes.
The take. Most LM tooling assumes Python, so a serious DSPy port to a different runtime is a signal that agent frameworks are becoming portable specifications rather than Python-specific libraries. The OTP angle is the real pitch: an agent that crashes and gets restarted by a supervisor, with its state intact, solves a reliability problem that Python agent frameworks usually bolt on with external orchestration. If teams running high-concurrency agent fleets in Elixir or Erlang shops adopt this over the next few months, it becomes evidence that agent frameworks are decoupling from any single language.
Definitions:
- DSPy: a framework for writing language-model calls as typed, measurable functions instead of hand-written prompts.
- BEAM: the Erlang virtual machine that Elixir runs on, built for concurrent, fault-tolerant processes.
- OTP (Open Telecom Platform): the set of Erlang/Elixir libraries for building supervised, restartable processes.
Cloudflare: launches cf, an agentic CLI covering the entire API
Cloudflare released cf in open beta on September 28, a new CLI generated from its OpenAPI schemas that covers more than 3,000 operations, compared to roughly 280 in its existing Wrangler tool. Cloudflare says agent use of Wrangler rose from single digits in 2025 to 25% in March 2026 to 48% last week, and that agents use nearly twice as many distinct commands per session as humans. cf defaults to JSON output for agents and pretty-printed output for humans, replaces Wrangler’s config with a new cloudflare.config.ts format for Workers, and ships with Vite as the default local dev server. It installs globally with npm i -g cf.
The take. Cloudflare is stating outright that it built a second CLI because agents had outgrown the first one, and the internal usage numbers back that up rather than reading as marketing. The JSON-default, config-as-TypeScript design is a bet that structured, LSP-readable interfaces beat human-legible tables even for the humans, since most people now review agent output rather than typing commands directly. The real test is whether agents actually prefer cf’s discovery mechanism over patterns already baked into their training data from years of Wrangler docs; if adoption climbs past the reported Wrangler agent-usage rate within two quarters, this becomes the reference design for agent-first CLIs.
Definitions:
- OpenAPI: a standard format for describing an API’s endpoints, used here to auto-generate CLI commands.
- LSP (Language Server Protocol): the protocol that gives code editors autocomplete and type checking, extended here to a config file.
Four different layers of the stack (documents, vision, agent frameworks, cloud infrastructure) are converging on the same design principle: build for the agent as the primary user, and let the human interface follow.
More drops at dropwatch.ai. Want them in your inbox? Subscribe below.