Generalizes Adaptive Computation Time from scalar halting on recurrent depth to budgeted control over typed deliberation graphs for retrieval, verification, tool use, and action.
Chaitanya Mishra Research Papers
Independent research on AI coding agents, durable execution, harness engineering, and systems design.
This archive collects long-form papers on coding-agent infrastructure, persistence architectures, runtime design, and evaluation methodology. Each paper is available as an indexable HTML page with a visible abstract, plus PDF and LaTeX source files for citation and reuse.
Argues that long-horizon AI agents need a typed decision state, not just transcript replay or more memory, and proposes commitment-carrying agent state as the missing primitive.
Introduces Worldline, a state-first harness architecture in which agents work against a causally versioned operational twin of the developer environment rather than a loose registry of tools.
Formalizes the runtime layer around coding agents as the Agent Harness Control Plane, an eight-function decomposition. Argues that for long-horizon software tasks, harness design choices often determine what an agent can see, what it may do, and whether it can recover from failure.
Develops a four-plane model of agent state and compares JSON, SQLite, and PostgreSQL as persistence backends for long-running coding agents. Compares Elixir, Erlang, Ruby, Python, TypeScript, Go, Rust, and Java as control-plane runtimes.
Argues that many agent stacks collapse interchange, checkpointing, effect history, and system of record into one mutable JSON artifact. Develops a persistence-role framework and operational-depth ladder, then compares JSON, SQLite, PostgreSQL, and BEAM-based runtimes.