Long story short: we needed something to actually do brownfield agentic engineering with: something that made it easier to reason about what an agent will see in context at any point in time, remove and flush context often, clear handoffs, better primitives.
Something we noticed is that LLMs have started co-evolving with their harnesses - anthropic gets more training data from you using claude code, claude code gets new features from claude first, etc etc - so we decided to change the base primitive from the llm call to the agent loop. almost any harness - codex, opencode, etc - is supported, you just need to generate a shim using one of the hanks like we do.
So we ended up building a runtime that executes sequenced AI programs - blocks of prompts and code that get loaded into context, run, monitored and then torn down.
We’ve been using it for a while, open-sourcing it in case its useful. Maybe it’s our visibility set but there’s not much out there focused on reliable executions and doing less so you can think more.
It’s single agent threaded (no parallel agent things), designed for headless (we’re a data company), no MCP or skills things - our general policy has been ‘build what you NEED’ and we just haven’t EVER needed those things.
hrishi•1h ago
Long story short: we needed something to actually do brownfield agentic engineering with: something that made it easier to reason about what an agent will see in context at any point in time, remove and flush context often, clear handoffs, better primitives.
Something we noticed is that LLMs have started co-evolving with their harnesses - anthropic gets more training data from you using claude code, claude code gets new features from claude first, etc etc - so we decided to change the base primitive from the llm call to the agent loop. almost any harness - codex, opencode, etc - is supported, you just need to generate a shim using one of the hanks like we do.
So we ended up building a runtime that executes sequenced AI programs - blocks of prompts and code that get loaded into context, run, monitored and then torn down.
We’ve been using it for a while, open-sourcing it in case its useful. Maybe it’s our visibility set but there’s not much out there focused on reliable executions and doing less so you can think more.
It’s single agent threaded (no parallel agent things), designed for headless (we’re a data company), no MCP or skills things - our general policy has been ‘build what you NEED’ and we just haven’t EVER needed those things.
Have fun!
Read more about the idea behind it here: https://southbridge.ai/hankweave