# Runtime Cut in Half and Prompts Trimmed by 80%: How Folding Subagents Into a Main Agent Loop Fixes Conway's Law in Enterprise AI

> Folding subagents into a main agent loop doubles runtime speed and cuts system prompts by over 80%. See how growth companies eliminate subagent sprawl.

- Published: 2026-10-25
- Author: Master Mind
- Canonical: https://aimaster.fi/en/artikkelit/ajoaika-puolittui-aliagenttien-yhdistaminen-ja-paasilmukka-arkkitehtuuri

Evaluating a multi-agent harness against a single agent loop cut execution runtime in half—delivering a 2x faster response—while progressive disclosure reduced Claude Code's system prompt length by over 80% (Anthropic, 2026). Meanwhile, 57% of enterprises now deploy AI agents for multi-stage workflows, and 16% run cross-functional processes across teams (Anthropic & Material, 2026). Folding subagents into a main agent loop replaces fragile department-by-department agent chains with one context-aware execution loop.

When every business unit builds its own specialist agent and a central team bolts an orchestrator on top, your software architecture becomes a mirror of your org chart. Software engineering knows this as Conway's Law. In AI deployments, Conway's Law turns a fifty-slide consulting diagram of boxes and arrows into a slow, expensive game of telephone.

## Why Does Conway's Law Break Multi-Agent Systems in Growth Companies?

Conway's Law breaks multi-agent systems because splitting agents by department or problem type forces constant context handoffs between isolated context windows. Every handoff drops critical details, multiplies token consumption, and adds seconds of latency (Anthropic, 2026). Instead of one coherent workflow, the company ends up maintaining dozens of overlapping subagents with conflicting prompts and no single owner for end-to-end quality.

In Anthropic's October 2026 enterprise engineering review, the Applied AI team identified five recurring failure modes when organizations assign one subagent per team or problem category:

- Lossy context handoffs (Telephone game): The coordinator passes a compressed summary to a specialist subagent, which loses edge-case constraints from earlier turns and returns another lossy summary.
- Subagent sprawl: Every department ships its own specialist agent without shared standards, leaving the company with dozens of overlapping bots and no clear ownership.
- The Frankenprompt: Every team appends its own rules to the shared instructions, creating contradictory directives and breaking brand voice across customer touchpoints.
- Localized, watered-down evaluations: Teams optimize metrics for their isolated subtask while nobody measures whether the customer's actual end-to-end problem was solved.
- Compounding latency and token tax: Every subagent invocation spins up a nested model loop that re-reads context and adds seconds of waiting time to the user experience.

Only 1% of Finnish companies use their data to create new business (Sitra), while 57% of Finnish SMEs use AI tools and 43% do not use them at all (2025). When a growth company's data is finally connected to AI, burying it under ten nested subagent handoffs destroys both speed and accuracy.

## How Does Folding Subagents Into a Main Agent Loop Improve Speed and Quality?

Folding subagents into a main agent loop improves speed and quality by keeping the entire task history, tool results, and plan inside one model and one context window. The model that writes the plan is the same model that executes it, allowing it to revise actions mid-turn as real data arrives. In Anthropic's enterprise comparisons (2026), a single agent equipped with modular skills consistently outperformed both monolithic prompts and multi-agent subagent harnesses.

Previously, teams split agents into subagents because stuffing every instruction and API schema into one prompt caused context bloat. Progressive disclosure eliminates that tradeoff. Instead of loading everything upfront, the main agent sees only compact descriptions of available [modular agent skills](https://aimaster.fi/en/artikkelit/100-500-kertaa-tiiviimpi-konteksti-tekoalyagentin-taitokirjasto-agent-skills) and tools, loading the full procedure only when the conversation requires it.

When Anthropic applied progressive disclosure of skills and tools to Claude Code for newer model generations, the system prompt shrank by over 80% (Anthropic, 2026). Once tool definitions in a system prompt exceed 30 to 50 tools, model tool-selection accuracy degrades; dynamic tool search with deferred loading keeps the active footprint lean while giving a single loop access to hundreds of capabilities.

|  |
|  |

Wait—let's examine the architectural comparison directly across the metrics that matter to a growth company's leadership team:

| Dimension | Fragmented Subagent Harness (Conway's Law) | Single Main Agent Loop + Skills & Tools |
| --- | --- | --- |
| Execution Latency | Slow; nested model loops add seconds on every handoff | Up to 2x faster runtime in production comparisons (Anthropic, 2026) |
| System Prompt Size | Bloated Frankenprompt or duplicated across dozens of agents | Over 80% smaller via progressive disclosure of skills and tools |
| Context Integrity | Lossy summaries between agents ('telephone game') | Full raw tool results and history preserved in one context window |
| Prompt Cache Hit Rate | Fragmented prefixes across isolated agent calls | 90–99% cache hit rate when skills load into the conversation prefix |
| Team Ownership | Each team owns a siloed subagent and local metrics | Central team owns main loop & end-to-end evals; teams own skills & MCP tools |

## When Should You Keep a Subagent vs. Fold It Into the Main Loop?

Fold a subagent into your main agent loop whenever specialists share context, run sequentially on the same data, or shape a unified customer conversation. Keep a separate subagent only when a task requires parallel execution in disposable context windows, when a single loop hits measurable degradation from massive intermediate data, or when strict regulatory isolation requires a separate compliance boundary (Anthropic, 2026).

A practical example from Anthropic's 2026 commerce architecture benchmarks illustrates the difference. A deep-research task—where an agent reads dozens of external documents, runs code, and explores dead ends—works well as a subagent called as a tool, because all intermediate noise stays isolated and only a compact answer returns to the main loop. Conversely, splitting order lookup, inventory checking, and shipping dates into separate subagents breaks reasoning. In Anthropic's support agent tests, separating order lookup from shipping caused the main agent to answer delivery questions on its own without consulting the shipping specialist in one out of two test runs.

When structuring the single main loop, apply Anthropic's one-third traffic rule (2026): instructions, safety guardrails, and core tools needed in at least one-third of all sessions belong directly in the global system prompt. Everything in the long tail moves into modular skills loaded as tool results so the global prompt cache stays warm—achieving 90–99% cache hit rates and 1.5x to 2x faster token reads at scale. For long-term user memory, writing facts asynchronously in a background thread adds zero latency to the main loop and improves fact recall by 13% compared to synchronous memory-saving tool calls (Anthropic, 2026).

## How Do Growth Companies Migrate From Siloed Bots to One Production Architecture?

Growth companies migrate from siloed bots to a unified main agent loop in stages: first by building an end-to-end task-completion evaluation suite, then by merging tightly coupled specialists that touch the same data, and finally by converting subagent prompts into modular skills and API calls into MCP tools. Every step is verified against real production test cases before retiring the old subagent.

AIMASTER is a Finnish AI partner that puts AI into production for growth companies — not as slide decks, but as working systems. We serve Finnish companies with revenue between €1M and €1B, and we do not serve listed companies. Custom AI solutions are delivered using an agile sprint model where one sprint equals 3 development days.

In our production work, we turn fragmented experiments into a unified architecture through three structured steps:

- [Master Plan](https://aimaster.fi/tuotteet/master-plan) is an AI strategy sprint that maps where AI generates the most value for your company — measured in euros. During this phase, we also identify where departmental silos would create unnecessary subagent chains and design a unified target architecture.
- [Master OS](https://aimaster.fi/tuotteet/master-layer) is the company's AI Operating System: a platform that safely connects your existing systems (CRM, ERP, documents), data, and agents for AI use. It enforces server-issued IDs, permission checks inside MCP tools, and high-hit-rate prompt caching.
- [Master Mind](https://aimaster.fi/tuotteet/master-mind) is the suite of AI agents running on top of Master OS that handles business processes autonomously. Instead of a fragile web of disconnected bots, Master Mind operates on shared context, progressive skill loading, and continuous [AI agent evaluations and QA](https://aimaster.fi/en/artikkelit/75-prosentin-tarkkuus-putoaa-42-prosenttiin-tekoalyagenttien-evaluointi-ja-laadunvarmistus).

We apply these same architectural principles across our public customer deliveries, such as digital marketing agency Tagomo and Jaajo Linnonmaa's Aini assistant, as well as our own portfolio products including Lyyli.ai and BookyAI. When the model that plans the work also executes the tools within a single disciplined loop, users get answers in seconds instead of waiting on a committee of bots.

If your company is planning multi-step AI workflows or untangling existing bot silos, start with a [Free Master Mind Analysis](https://aimaster.fi/analyysi). We will map your highest-value processes and show how a unified agent architecture works in production within 3-day sprints.

## Frequently asked questions

### When should you fold a subagent into the main agent loop?

You should fold a subagent into your main agent loop whenever two specialists touch the same data, chain sequentially on each other's output, or require the user's conversation history. In Anthropic's 2026 enterprise evaluations, moving from an orchestrator-subagent harness to a single agent loop cut execution runtime in half (2x faster) while improving answer consistency.

### When does a separate subagent still make sense in production?

Keep a separate subagent only in three cases: when a single loop suffers measurable accuracy degradation from context bloat, when independent subtasks must run in parallel with disposable context windows (such as deep document research), or when a separate domain requires strict isolated compliance boundaries.

### How does progressive disclosure prevent system prompt bloat in a single agent loop?

Progressive disclosure loads only short descriptions of skills and tools into the initial context and fetches full instructions only when the agent needs them. Anthropic used progressive disclosure to reduce Claude Code's system prompt by over 80% (2026), eliminating the context bloat that previously forced teams to split work across subagents.

### Should domain instructions live in the system prompt or in agent skills?

Use the one-third traffic rule from Anthropic's 2026 architecture benchmarks: instructions, safety rules, and core schemas required in at least one-third of all sessions belong in the global system prompt. Procedures needed less frequently belong in modular skills loaded as tool results so they stay inside the prompt cache prefix.

### How can multiple teams collaborate on one main agent loop without breaking production?

Multiple teams contribute to a single agent by owning domain-specific skills, MCP tools, and test cases rather than separate subagents. One small central team owns the shared system prompt and end-to-end evaluation suite, gating every new skill or tool against golden task-completion tests before deployment.