Type something to search...
to navigateto selectESC to close

Loosen the Leash

A build log about earning autonomy for coding agents one notch at a time: a harness, a memory and a factory loop, each refusing the other two's job. A human merges.

Start from the beginning

The layers alone

4 parts
draft

Three layers, three refusals

Series opener: the harness guides, the memory remembers, the factory decides, and a human merges. Each layer refuses the other two's job, and that is what makes them safe to stack.

draft

The harness guides

Part 1: a definition of done the agent cannot see, a budget that holds, and a request for safety that refuses rather than pretends. Everything around the model, declared.

draft

The memory remembers

Part 2: memory that lives in git beside the code, keeps private notes private, fits a budget, turns a repeated failure into a lesson, and only ever advises.

draft

The factory decides

Part 3: a loop that turns a labelled ticket into a draft pull request with no model in it. Every decision logged, restarts that come back clean, events treated as hints, and policy you can replay.

The layers together

3 parts
draft

The factory drives the harness

Part 4: put the harness inside the factory and point it at a tracker anyone can write to. A lane that can only tighten, a pinned harness, an agent with no keys, and a ticket that gives orders.

draft

The factory remembers, and memory never decides

Part 5: a factory writes every failure to memory with its exit code and its own words, lets counters make the decisions, and keeps the line moving when memory is down or was never installed.

draft

Together, then taken apart

Part 6: one ticket through the harness, the memory and the factory to a merged pull request, then each layer removed in turn to see what the other two still do.