Thu Oct 22

October 22, 2026

This is the Bronco dev log for Thursday, October 22, 2026! It's time to huddle. 🐴 This is written by Robert Austin.

  1. Skill-based Context Management
  2. Project Re-do
  3. Bronco reads your files in the background
  4. Optical Character Recognition
  5. Fixture joins the Map

1. Skill-based Context Management

Before, a Bronco session carried 60 tools in context at once, and the instructions overloaded the model. After, a session starts with only use_skill, skill_search, and todo; it loads a skill from disk, and the skill presents its member tools. Skill search uses the knowledge base's semantic-lexical hybrid search to find skills by meaning.
Before, a session carried 60 tools at once. After, it starts with use_skill, skill_search, and todo, and gains a skill's tools when it loads that skill.

Bronco has 'tools' which it uses on your behalf. It has 60 today and that's too many for Bronco. That's because the tools have to be explained to Bronco's Large Language Model (LLM) and the more instructions it gets, the more difficulty it has in focusing and making decisions.

Originally I worked around this issue by organizing Bronco's tools into discrete agents. Unfortunately this works extremely well when you design the agents for a specific problem but otherwise each agent will likely have too little context to understand its role or the overall progress of work and decision-making ends up being a coordinated effort. The LLM itself is maybe the correct layer in the stack to fix that problem.

Today I introduced a new 'skill' based design for tool discovery. Now Bronco has a set of active 'skills' at any time. Each skill is a file listing a set of instructions and the tools required to perform the task. Each Bronco session receives instructions for any tools it needs to complete its tasks.

Bronco sessions now start off with only the tools required to find and use skills and to organize todo-lists. Bronco starts by considering the user's request and then searches for instructions on how to solve the request. Skills vary in how specific and general they are and the agent can use several at once. Basic skills help the agent browse the Web while advanced skills help the agent perform domain-specific tasks like helping with groceries, photo editing, or spreadsheets.

Skill search uses the knowledge-base, a semantic-lexical-hybrid search engine, which allows Bronco to search for skills by meaning or content. Bronco is nudged to find new skills after each user message. Once the basic version of this is working, I intend to streamline and optimize the approach by reducing LLM turns required to get fresh skills.

2. Project Re-do

A Project is a folder plus the data Bronco keeps about it. Two Projects sit side by side, ~/Bronco, the default, and ~/Work. Each lists its sessions and keeps messages, tool results, and generation settings. Each has its own background reading switch: paused for one, running for the other. Sessions are grouped with the files they read and write, search is scoped to the folder with its own indexing settings, and a Project can override text Bronco injects into the model's instructions. The default Project is ~/Bronco, and creating or deleting a Project never deletes the folder.
A Project is a folder plus the data Bronco keeps about it: its sessions, its background reading switch, its scoped search, and its overrides to the text Bronco injects into the model's instructions.

Projects start as folders on your computer. But once you tell Bronco about one, it becomes searchable, and Bronco can read and work on the files with you.

To search your files, create a new project in Bronco. Bronco will immediately begin reading your files.

3. Bronco reads your files in the background

Then, a reply sat behind five batches of background file reading, so nothing came back until the reading stopped. Now, the waiting reply is picked first, and reading resumes only when no reply is waiting. A batch finishes its unit, then the next pick is a reply.
Before, a reply waited behind five batches of background file reading. After, the waiting reply is picked first and reading resumes when no reply is waiting.

I was giving a friend a demo of Bronco the other day and it just wouldn't work. Nothing I typed got any replies. Finally at the end of the call I looked back and Bronco had answered. Annoyed I looked into the issue right away and found that Bronco was prioritizing work on reading the users files in the background over replying directly. Going forward, Bronco will always reply to you before reading files in the background.

4. Optical Character Recognition

A scanned handwritten journal page with wrapped lines, an aside figure, a margin note, and a page number. Beside it, the same page as clean structured text with a heading, paragraphs, a figure placeholder and caption, and a page number. An OCR model reads the layout, not just the letters. Added today with initial tests; OCR for PDFs whose text lives inside images is next. It takes less reading, fewer turns, and less energy than asking the vision model to transcribe.
An OCR model reads a scanned page's layout, not just its letters, and returns structured text. It takes fewer turns and less energy than asking the vision model to transcribe.

Bronco can analyze images. And it can read the text from an image. It can read the serial number off of a product in an image, for example. But if you were to ask Bronco to convert a stack of scanned handwritten journals into a text document, it would struggle. It would have to, essentially, read and then transcribe. Bronco isn't an expert at knowing what to do in complex document processing so it would struggle knowing how to treat lines and text that wrap, are ordered unusually, have aside-figures, or any number of other complex page layouts. But an OCR (Optical Character Recognition) model can help. By adding a new special model, Bronco will be able perform better and much faster and with less energy when processing images containing scanned documents.

Today I have added support for an OCR model to Bronco and have begun initial tests. In the near future I hope to support automatically applying OCR to PDFs when they contain text embedded into images.

5. Fixture joins the Map

The interactive map with Fixture selected. Its colorful logo sits on the top cube of the Agents layer, the company index on the left highlights Fixture, and the inspector panel reads Fixture, Agents, with a description of its agentic CRM work.
The map with Fixture selected: its logo on the top cube of the Agents layer, the company index highlighting Fixture, and the inspector reading Fixture, Agents.

I added Fixture to the industry map. It's an agentic CRM: it reads email, calls, and Slack, updates deal records, prepares meeting briefs, and drafts follow-ups without being asked. It sits in the Agents layer with the other agent tools. See the map post.