Subagents

Cursor cut agent token cost by shrinking static instructions, loading tools on demand, improving cache boundaries, and using subagents more selectively.

Reliable agent skills need more than prompts: use hooks as automatic guardrails, test across models, and compile harness-specific variants when instruction-following differs.

Coding agents often report reviews as complete despite unread files. Treat final messages as untrusted summaries and verify coverage, commands, and artifacts from the execution trace.

Delegating retrieval to a trained search subagent can keep noisy results out of the main agent’s context while cutting the token, latency, and cost burden of search-heavy workflows.

Knowledge agents need intent decomposition, specialist researchers, and multiple retrieval primitives; copying a coding-agent loop fails when evidence is ambiguous and context-dependent.

RLMs keep large inputs in a programmable environment, letting a model inspect them with code and delegate subsets to submodels instead of attending to every token at once.

BATON explores robot subtasks independently, stores their solutions, and manages handoffs between them, replacing multiplicative whole-task search with additive composition.

A walkthrough of Anthropic’s architect exam doubles as a harness checklist: inspect stop reasons, narrow subagent scope, isolate outputs, and compact long-running context.

The talk argues that agents are an early interface, not the endpoint, and demonstrates reusable primitives for AI-native software: persistent subagents, compressed context, loaded skills, generated UI, and voice.

Production agent velocity depends less on model swaps than on automating the developer loop: isolate parallel work, encode workflows as skills, connect tools, and keep humans as verifiers.

A fixed chain of specialist agents lost context and produced incoherent recommendations. The replacement separates deterministic detection, gives one agent end-to-end ownership, and uses subagents only for bounded investigations.

Vercel AI Gateway added Gemini 3.6 Flash for coding and web work and Gemini 3.5 Flash-Lite for scoped subagent tasks, both available through AI SDK model IDs.

Cursor’s SQLite experiment suggests swarm harness design can matter more than model mix: separate planner and worker contexts reduced coordination churn while cheaper workers handled most tokens.

OpenCode provides open-source terminal and desktop coding agents, with separate full-access and read-only modes plus a general-purpose subagent.

career-ops shows how a coding-agent harness can run personal operations: structured scoring, parallel workers, document generation, and human approval around a durable data store.

PI-Desktop packages coding agents, reviews, permissions, models, MCP, skills, and subagents into a local-first workspace, but its early-preview interfaces remain unstable.

A reusable Claude Code harness bundles subagents, skills, hooks, rules, evals, and context persistence. Treat it as a pattern library to adapt, not a configuration to enable wholesale.

Muse Spark 1.1 gives agent builders a 1M-token multimodal model with parallel tool calls, MCP support, and the option to run as a main agent or subagent.