Amazon Bedrock AgentCore Launches Automated AI Agent Optimization via Trace-Based Reflector System
Summary
Updated Sep 18 — Amazon launched a new AgentCore runtime with elastic memory reclamation, consistent cold starts, and billing tied to actual resource consumption.
- • Amazon Bedrock AgentCore automates agent system prompt optimization using production traces
- • Sub-Agent Reflector achieves 95.83% on AppWorld benchmark, 23 points above baseline
- • Single Agent Reflector delivers competitive quality in 6 minutes vs. 108 minutes for GEPA
- • Guardrails prevent optimization drift: length ceiling, safety check, no verbatim trace phrases
Updates
New AgentCore runtime launched: elastic memory, consistent cold starts, usage-based billing
Amazon announced a redesigned AgentCore runtime purpose-built for production agents. Key improvements: dynamic memory management that reclaims memory as sessions release it (ending peak-hold billing), consistent cold start times regardless of container size or concurrency, and billing that tracks actual resource consumption rather than peak provisioning.
Memory reclaimed at session end; no longer billed at peak allocation for full duration
The previous runtime held memory at peak allocation for the full session duration, causing long-running or bursty agents to overpay. The new design reclaims memory as soon as it is released by the agent — eliminating the cost penalty for ambient and event-driven workloads.
Runtime extended for ambient always-on agents triggered by events, not just sessions
The redesigned runtime extends AgentCore's serverless model from short interactive sessions to long-running, ambient, and event-triggered agents that run unattended and are increasingly launched by other agents — addressing the shift from chat-style to autonomous production workloads.
Thousands of production teams running agents on AgentCore runtime since original launch
Amazon cited thousands of development teams using AgentCore runtime for production workloads, with feedback identifying three priorities: faster responsiveness at scale, finer resource control, and cost precision — all addressed by the new runtime design.
Details
AgentCore optimization automates agent system prompt improvement from production traces
Amazon Bedrock AgentCore's optimization capability replaces the manual process of trace review and prompt tuning. An agentic reflector analyzes evaluated trace sets, identifies success/failure patterns, and proposes configuration changes validated through offline evaluation and live A/B testing.
Reflector agent uses shell tool access to full trace filesystem rather than summarizing
Rather than truncating or summarizing traces to fit a context window, the reflector accesses the full trace corpus via a filesystem using shell tools (grep, cat, diff, jq) to explore and compare traces, then proposes targeted edits to agent configuration.
Single Agent Reflector: 81.55% on AppWorld in 6 minutes, 18x faster than GEPA at comparable quality
On AppWorld benchmark (baseline 72.62%), the Single Agent Reflector achieves 81.55% in 6 minutes and 20 turns, vs. GEPA's 79.63% in 108 minutes and MIPROv2's 74.40% in 216 minutes — best efficiency-to-quality trade-off of the methods tested.
Sub-Agent Reflector achieves 95.83% on AppWorld, 23-point lift and 16 points above next-best
The experimental Sub-Agent Reflector deploys independent sub-agents to analyze individual traces at three levels (surface, turn-level, cognitive), then aggregates findings. Achieves 95.83% on AppWorld and 79.15% on WebShop — highest of all four methods tested.
Guardrails block common optimization drift: 20% length ceiling, safety screen, no verbatim phrases
Every candidate configuration update must pass three guardrails: a 20% length growth ceiling (rejected and re-requested if exceeded), safety standards screening before promotion, and a prohibition on verbatim phrases from optimization traces to prevent surface-level overfitting.
Sub-Agent Reflector released as experimental open-source in Strands GitHub repository
The higher-quality Sub-Agent Reflector is available as an experimental feature in the Strands open-source GitHub repository. The Single Agent Reflector and managed optimization workflow are available in the Amazon Bedrock AgentCore console.
Source: AWS Machine Learning Blog (September 16, 2026)
What This Means
Amazon is automating one of the most manual and time-consuming parts of production AI agent development: iterative system prompt optimization. By using an agentic reflector with filesystem access to full trace corpora rather than a fixed summarization pipeline, AgentCore can identify failure patterns that standard approaches miss. The benchmark results are strong — particularly the Sub-Agent Reflector's 95.83% on AppWorld — though these are Amazon's own evaluations and independent validation remains important. For enterprise teams running production agents on Bedrock, this meaningfully reduces the iteration cycle from days of manual trace review to minutes of automated optimization.