saved
Planning with Agents: Divided Worlds, Boundary Objects, and Thicker Interfaces
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
Maggie Appleton of GitHub Next argues that text-only agent plan modes are a poor medium for planning complex software work with agents. Drawing on Lucy Suchman's Plans and Situated Actions and Star and Griesemer's boundary objects, she says humans and agents do not need shared understanding. They need thick, visual, interactive and social interfaces that agents build to make human decisions easier.
ideas
- Plans are a weak resource. Using Gladwin's contrast of European and Chuukese navigators via Suchman, she argues a plan is an imagined journey that orients action but cannot predict what happens on the way.
- Walls of text flatten thinking. CLI plan modes reduce decisions to multiple choice text, cutting out the visual, embodied and social ways people actually reason about hard problems.
- Boundary objects, not shared worlds. Humans and agents live in divided worlds, so the goal is artifacts that serve each side while keeping a consistent identity, like the specimens and maps of Berkeley's zoology museum.
- Thicker interfaces built by agents. Treating agent labour as functionally free, agents can build live shadow and easing editors, state machine explorers, and branch-by-branch prototypes with interactive reports, as in a retry-logic demo where one option clearly wins.
- Chopin as the testbed. She is exploring these ideas in Chopin, a multiplayer, real-time MDX planning prototype built with Krzysztof Cieślak at GitHub Next.
quotes
“Plans are best viewed as a weak resource for ad hoc activity.”
“Plans orient action. They cannot predict the whole journey.”
“Humans and agents do not need shared understanding. They need good boundary objects and thick interfaces.”