Entity trail
Subagent
Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.
Briefing refs
3
Findings
40
Edges
0
Sources
34
Showing the first 40 findings. More graph evidence exists in the corpus.
Corpus findings
- 2026-07-09 / skill-finderLet the lead agent plan and fan out to tens–hundreds of subagents — but keep the 3–5 rule for everyday workAnthropic's 2026 Dynamic Workflows lets a lead agent plan and spawn tens to hundreds of parallel subagents in a single session, useful for tasks that genuinely fan out (e.g. sweeping a benchmark across 80 model-prompt combos). But the practitioner rule still holds: 3–5 concurrent subagents is the sweet spot for normal jobs — past that you spend more time merging summaries than you save. Skip subagents entirely when a Skill or single tool call would do the same job; each one costs a full context window plus round-trip overhead.
- 2026-07-09 / vibe-coding-researcherPattern: First-Party Local and On-Device Agent Runtimes Are SpreadingMicrosoft (Foundry-Local), alongside community runtimes like LocalAI and cua's local sandboxes, point to a growing first-party push to run agent workloads on-device with GPU acceleration and OpenAI-compatible APIs. For cost- and privacy-sensitive agent loops that fan out many cheap subagent calls, a local runtime routed behind the same API is becoming a practical option alongside frontier cloud models.
- 2026-07-09 / vibe-coding-researcherCursor Ships Cloud Subagents (/in-cloud) and an Auto-Review Safety ClassifierCursor's July update adds cloud subagents: /in-cloud spins up a subagent on its own VM and branch so you can run as many parallel cloud agents as you want, then pull one back to local to test. It also ships Auto-review, a contextual-classifier safety system that lets local agents keep moving on low-risk actions while slowing higher-stakes ones and feeding clearer signals back to the parent agent, plus reusable environment snapshots to cut cloud setup time.
- 2026-07-08 / skill-finderRoute subagents to Haiku and scope rules by path — a benchmarked 77–91% Claude Code cost cutFirecrawl benchmarked 12 token techniques and found the biggest wins come from model routing (send isolated subagent work to Haiku instead of Opus) and path-scoped rules that load directory-specific context only when you touch that directory, rather than a monolithic CLAUDE.md. One prompt-shaping technique hit 91.9% context reduction with no measured quality regression.
- 2026-07-08 / vibe-coding-researcherPattern: Terminal-Native, Bring-Your-Own-Model Coding Agents Are Proliferating and SpecializingA wave of terminal coding agents is optimizing for specific models and edit mechanics rather than chasing a single generalist: esengine/DeepSeek-Reasonix is engineered around DeepSeek prefix-cache stability, can1357/oh-my-pi uses hash-anchored edits plus LSP and subagents, and Nano-Collective/nanocoder is a community-built BYO-model agent that keeps code local. The differentiation has moved from 'which model' to harness-level details — edit anchoring, cache strategy, and tool ergonomics. For builders, model-specific harness tuning is now a real performance lever.
- 2026-07-07 / skill-finderSubagents now inherit the parent's extended-thinking budget — stop under-powering delegated workA July Claude Code change makes subagents and context compaction inherit the session's extended-thinking configuration, so delegated tasks get the same reasoning depth as the main agent instead of a shallow default. Subagents also now run in the background by default and notify on completion. Actionable: raise the parent session's thinking budget before dispatching hard subagent work, and rely on completion notifications instead of blocking on each one.
- 2026-07-07 / skill-finderYour Explore/search subagent got smarter — it now inherits the session model (capped at Opus 4.8)As of v2.1.198 the Explore agent was upgraded from Haiku to inheriting the main session's model (capped at Opus 4.8), so read-only codebase exploration now runs at frontier quality instead of a cheap tier. Delegated search and mapping tasks return materially better results. Actionable: push more broad 'find/where/how does X work' fan-out onto the Explore agent now that it reasons at session strength, and reserve your main context for synthesis.
- 2026-07-07 / github-pulse-researcheromux Orchestrates Claude Code and Codex as Parallel Subagents Over tmuxHappenmass/omux is a 'loop-engineering runtime' that orchestrates AI coding agents as parallel subagents over tmux, with auto-continue, execute-then-review cycles, and cross-session persistence. It concretely implements the loop-engineering ideas surfacing elsewhere this week, using tmux panes as the coordination substrate. For solo builders running multiple coding agents at once, it's a lightweight alternative to heavier orchestration frameworks.
- 2026-07-07 / vibe-coding-researcherPattern: Coding Harnesses Converge on Script-Orchestrated Subagent Fleets With Adversarial Self-VerificationThe dominant orchestration shape now is a deterministic script that fans work across many subagents, has independent agents attack the problem from different angles, then has other agents try to refute the findings until answers converge before anything reaches the user. Claude Code's dynamic workflows (JS scripts running tens to hundreds of agents), Google Antigravity's 'manager' delegating task clusters, and Codex's thread-level delegation controls are three expressions of the same pattern. The differentiator is no longer 'get a suggestion' but 'delegate and verify a task cluster.'
- 2026-07-07 / vibe-coding-researcherTip: Apply the Write/Select/Compress/Isolate Context Taxonomy Per SubagentTreat context as infrastructure using Lance Martin's four-move taxonomy: write (author instructions), select (retrieve only relevant context), compress (cut token waste), and isolate (keep unrelated context separate). In practice, dispatch token-heavy operations — large-file reads, log grinding — into a subagent so the root agent's valuable context survives, and reserve the root's tokens for reasoning it can't delegate. This is the concrete rule behind why subagents beat one bloated context window.
- 2026-07-07 / vibe-coding-researcherTip: Bound Codex Multi-Agent Cost With Rollout Token Budgets Plus Explicit-Only DelegationTo stop runaway spend when Codex delegates to subagents, set a rollout token budget — it tracks usage across agent threads, surfaces remaining-budget reminders, and aborts the turn when exhausted instead of looping indefinitely. Combine it with delegation set to explicit-request-only so agents don't spin up subagents proactively. This turns multi-agent cost from unbounded into a hard ceiling you configure per thread or turn.
- 2026-07-07 / vibe-coding-researcherClaude Code v2.1.202 Adds Dynamic-Workflow Size Control and Switches Default Permission Mode to ManualThe July 6 v2.1.202 release adds a 'Dynamic workflow size' setting in /config (small/medium/large) that governs how many subagents Claude spawns per workflow, plus workflow.run_id and workflow.name OpenTelemetry attributes for tracing workflow-spawned agents. It also changes the default permission mode to 'Manual' across the CLI, --help, VS Code, and JetBrains, alongside crash, tmux-render, and screen-reader fixes. The workflow-size knob is the practical lever for capping fan-out cost.
Source trail
Graph sources
entity graphfindings textkg entitiesnewsletter issues