AWS Strands Harness: Same Model, 77% Lower Cost – Why the Agent Runtime Now Matters More Than the Model

Benchmark: Same Model, Vastly Different Cost and Accuracy

AWS ran a harness benchmark across six suites (ALFWorld, ContextBench, GAIA, WebShop, τ³-bench, Terminal-Bench 2.1) using EC2 distributed infrastructure. All harnesses used the identical underlying model: Claude Fable 5 . Over 89 trials, Strands Harness cost $56.29 and scored 69.7 ; Claude Code cost $248.05 and scored 61.8 . Other harnesses fell in between: oh-my-pi ($86.83, 69.7), OpenCode ($73.42, 66.3), DeepSeek Harness ($40.30, 59.5). Strands achieved ~28% lower average cost than other harnesses across all six benchmarks, with Terminal-Bench showing the widest gap.

How Strands Compresses Context: Two Key Rules

1. Tool Result Offloading at ~1,500 Tokens

Long-running agents bloat context mainly through tool outputs (shell logs, search results, test reports). Strands' ContextManager truncates any tool result exceeding ~1,500 tokens , keeping only a head/tail preview in the active context. The full output is written to a Stash and can be retrieved later via retrieve_context. This is not deletion – it's tiered storage: hot context stays small, cold data remains accessible.

2. Compaction at 85% Context Utilization

When the context window reaches 85% capacity , Strands automatically summarizes older messages while preserving the most recent 4 messages . If compaction still leaves overflow, a ContextRecovery mechanism continues the agent loop instead of aborting with a context-length error. The article illustrates the compounding effect: a tool returning 4,000 tokens over 20 rounds would naively accumulate 80,000 tokens; offloading removes repeated re-reading costs across subsequent model calls.

Layered Memory: Context Window, Session, Long-Term Memory

Strands separates three storage tiers:

Context Window – working memory, aggressively managed via offloading and compaction.

Session – full conversation state persisted to ./.agent/sessions; restores task progress after process restart using the same Session ID.

Long-Term Memory – managed by MemoryManager with three capabilities: Recall (search history), Injection (prepend relevant memories to prompt), and Extraction (write new durable knowledge). Recall and Injection are enabled by default; Extraction requires explicit opt-in.

This mirrors traditional storage hierarchy: context ≈ RAM, session ≈ task state, memory store ≈ persistent knowledge.

Skills with Progressive Disclosure

Instead of stuffing all skill instructions into the system prompt, Strands loads only skill names and descriptions initially. Full instructions are fetched via tool call only when the model decides they are needed . This mirrors the tool-result offloading philosophy: don't pre-load speculative context.

Prompt Caching and the Real Cost Equation

Strands enables caching: auto by default. On providers with explicit cache control (Bedrock, Anthropic), system prompts, tool definitions, and stable message prefixes become reusable cached blocks. The article presents the effective cost formula:

Effective Context × Call Count – Cache Hits – Offloaded History + Compaction Cost + Subagent/Tool Call Cost

Thus, identical model pricing can yield wildly different bills depending on harness implementation – motivating harness-level benchmarks.

Model × Harness × Workload: The New Evaluation Target

Strands standardizes a Model Interface across providers (Bedrock, Anthropic, OpenAI, Google, Ollama, LiteLLM, Mistral, Llama API, llama.cpp, Vercel). The same harness can swap models without rewriting shell, memory, skills, MCP, session, or context logic. It also decouples the harness from cloud platform (runnable on Cloudflare Containers, Azure Container Apps, Google Cloud Run, Amazon ECS, Bedrock AgentCore, Modal).

However, provider differences remain: prompt caching support, tool schema handling, reasoning effort, context window, parallel tool calls. The benchmark proves Model × Harness is a coupled pair – the harness controls what the model sees and when it acts; the model determines how well it uses that information. Therefore, production evaluation must test Model × Harness × Workload triples: cost, success rate, step count on real tasks.

Conclusion: The "Body" Is Stabilizing While the "Brain" Keeps Swapping

AWS's release signals a shift: agent scaling is moving from model improvements to the surrounding runtime system. The model sets the capability ceiling, but context management, memory, tools, skills, cache, and execution loops determine how much of that ceiling is realized. Strands aims to fix the "body" (harness) so developers only swap the "brain" (model). Caveats: data is AWS self-reported; full paper not yet public; DeepSeek Harness shows lower cost ($40.30) but lower accuracy (59.5), illustrating the trade-off.

References:

Strands Agents, "Introducing Strands harness: frontier performance with 28% lower token cost", Sep 21, 2026.

Strands Agents Documentation – harness, configuration, context management, memory, skills.

The New Stack, "AWS open-sources an AI agent it says is 45% cheaper than Claude Code and Codex", Sep 21, 2026.