Experiment · EXP-011

Should a coding agent live in one endless chat?

Long sessions versus a new chat per task, with bugs written as tasks.

Long threads mixed stale errors, old plans, and finished work. Starting a new chat per task, and deleting outdated tasks, reduced confusion in practice. Related to the token-waste work.

Hypothesis: One continuous thread preserves the most context and therefore performs best.

Research question

Does keeping every coding task inside one continuous agent conversation improve performance because more history remains available?

Observed behaviour

Long threads accumulated stale errors, superseded implementation plans, finished tasks, old assumptions, and repeated explanations. The conversation contained more information, but the ratio of relevant information to total information became worse.

Alternative

I moved toward a new conversation per meaningful task while keeping durable project knowledge in rules, documentation, and the codebase itself. Bugs were written as explicit tasks. Outdated task descriptions were removed or replaced rather than left as competing instructions.

Result

Task-scoped conversations reduced confusion in practice. They also reinforced a broader conclusion: preserving more context is not the same thing as preserving the right context.

Finding

The infinite thread preserves volume. It does not preserve relevance.

Abubakar, Imam. "Should a coding agent live in one endless chat?." Imam Abubakar AI Research, 2025. https://imamabubakar.com/research/exp-011-one-infinite-thread

This archive contains independent research, experiments, architecture studies, and technical observations. Some findings are exploratory and should not be interpreted as peer-reviewed scientific conclusions unless explicitly stated. Architecture studies of third-party products are based on publicly observable behaviour, published technical information, and engineering analysis. They are not claims of access to private source code or infrastructure.