Research

Reviewed on: 2026-06-28.

A loop needs a place to keep work that outlives a context window, and a set of gates that bound what an autonomous run can break. GitHub Issues and Projects supply the first…

A worktree is a second working directory attached to the same repository, with its own checked-out branch and its own index, sharing one object database and one history. For…

Loop engineering is the practice of designing the control system that prompts, verifies, and stops a coding agent, instead of being the person who types the next prompt. Addy…

Context observability explains how a model request was assembled and how that request behaved. It can cover candidate context, selected context, transformations, provider…

Token optimization in tool-using agents covers several distinct problems: reducing verbose inputs, limiting fixed prompt overhead, removing context that has outlived its…

Tau names two related coding-agent efforts: dpc's Tau emphasizes Unix-native process composition, while Hugging Face Tau is a small Python teaching harness whose core lesson is…

OpenCode Zen is an optional OpenCode AI gateway that exposes a curated model catalog through direct API endpoints, including Big Pickle as a free OpenAI-compatible…

Time to first token is the user-visible latency between submitting an LLM turn and receiving the first generated output token or stream event.

ACP standardizes the boundary between coding-agent clients and coding-agent implementations: clients own the user interface and environment, agents own the model loop, and both…

iocraft is a Rust terminal UI library that uses a declarative, React-inspired component model with hooks, flexbox layout, terminal events, and mock-terminal testing support.

Local agent memory can combine inspectable source files with rebuildable SQLite indexes: FTS5 for lexical retrieval and sqlite-vec for semantic vector recall.

MCP defines a JSON-RPC lifecycle for connecting clients to external tool and context servers, with explicit initialization, capability negotiation, transport rules, tool…

Provider APIs expose similar agent concepts through different wire shapes, so a coding harness needs provider-specific adapters that normalize messages, reasoning, tool calls…

SafeExec is a Bash-based command interception layer for local human and AI-agent shell sessions: it gates a small set of destructive command shapes with a real TTY confirmation…

OpenAI presents ChatGPT subscription access for Codex as an OAuth/access-token credential path for Codex clients, distinct from OpenAI Platform API-key access.

A strong context-control system treats model context as an inspectable, user-controlled working set assembled from durable session state, scoped project instructions, explicit…

Terminal apps trigger desktop notifications through one of three paths: directly calling the platform notification API, shelling out to a platform helper, or emitting a…

Good internals documentation gives readers a stable mental model and gives contributors a safe path into the codebase without turning the public docs into a source listing.

OpenAI's harness engineering essay argues that agent-first software development shifts engineering effort away from hand-writing code and toward designing the repositories…

Notes on Letta's memory-first agent model, memory blocks, archival memory, shared state, MemFS, dreaming, and stateful agent APIs.

Notes on Polytoken's local-first coding-agent daemon, sessions, facets, prompt input, work management, and permissions.

Modern terminal coding-agent UIs converge on a chat transcript, bottom prompt, visible execution state, commands, review loops, and temporary detail surfaces; the main design…

OpenAI's Codex guide argues that Codex performance depends as much on harness prompt/tool design as on model choice: prompts should emphasize autonomy, efficient exploration…

These crates fit different layers of text handling: unicode-segmentation teaches correct edit boundaries, ropey teaches when a real editor buffer is worth its cost, and…

An Agent Skill is a directory-level instruction package whose required SKILL.md frontmatter makes it discoverable, while optional scripts, references, and assets are loaded…

The Piebald AI repository presents Claude Code as a prompt-and-tool system made of many small runtime-interpolated fragments for agents, tools, modes, classifiers, reviews…

Codex's Rust prompt crate renders specialized prompt text from typed runtime state, separating permissions, review, goals, compaction, realtime, and patch-tool instructions.

Goose keeps its prompt surface as small markdown templates for agent identity, planning, compaction, subagents, recipes, app generation, permissions, and tiny-model behavior.

Gridland uses Yoga as the central geometry engine in a React-style terminal UI runtime: JSX elements become renderables, renderables own Yoga nodes, React tree mutations update…

OpenCode Go is a paid OpenCode provider that curates open coding models behind OpenAI-compatible and Anthropic-compatible API endpoints, with model IDs, usage limits, and…

Pi's .pi/prompts directory is a compact set of procedural slash-command prompts for maintainer workflows, written as explicit checklists with safety rules and expected output.

Prompt libraries can inform input editing, completion, history, and multiline behavior, but they should not automatically own a coding-agent terminal renderer. A rich agent UI…

Yoga is an embeddable, cross-language layout engine that computes the size and position of boxes using a CSS/Flexbox-like model; it does not render pixels, terminal cells…

Public coding-agent prompts converge on layered prompt assembly: base identity, behavior policy, tool/editing constraints, environment and project context, then the user's…

There is no single standard coding-agent session format. Mature agents converge on a few concepts that survive specific implementations: durable append-only records, explicit…

Pi argues that a coding harness can stay small and inspectable by exposing simple tools, explicit context, event-streamed agent state, file-based planning, and terminal-native…

AGENTS.md should be treated as scoped, repo-owned operating instructions: useful enough to load automatically, but never powerful enough to override user instructions, tool…

Agent harnesses benefit from treating filesystem traversal and text processing as typed, bounded operations. fd and ripgrep are strong discovery/search primitives; sed and awk…

Gridland's AI chat interface is a compact two-panel terminal-style app: fixed conversation sidebar, flexible message stream, prompt input with command/file suggestions…

Herdr is a Rust/Ratatui terminal workspace manager for coding agents: it keeps real terminal processes in panes, adds workspace/tab structure, tracks agent state, and exposes a…

Ratatui leaves application architecture to the app. The durable lesson is that terminal agent apps stay simpler when state, messages, update logic, terminal I/O, and rendering…

Release scope for a terminal coding harness should be defined by user-visible contracts: provider reliability, sessions, configuration, file-change safety, packaging, docs, and…

Ratatui UIs can be tested by rendering into a fixed-size TestBackend and snapshotting the backend buffer with insta.

A good first pass on a new codebase combines two maps: Git history shows where people have been working, struggling, and firefighting, while a structural code map shows which…