MoEngage turned roughly 400 private APIs into permission-scoped agent tools with one managed gateway, and kept a 20,000-token tool catalog out of the agent's context
AspireAI community · Amazon Web Services
The hard part of connecting agents to a company's existing APIs is rarely the calling. It is that a marketing platform like MoEngage has roughly 400 internal REST APIs built by 25 teams across seven AWS Regions, all VPC-private for data-residency reasons, and an agent catalogue that large will not fit in a context window. MoEngage's answer was to give each product team one job - maintain an OpenAPI spec and mark which operations should be tools - and hand the protocol layer to a single platform team.
Amazon Bedrock AgentCore Gateway, a fully managed MCP server, turns those OpenAPI specs into tools and evaluates Cedar policies server-side before the tool list is built, so an agent that should not see write operations never receives them. Identity travels in headers the model cannot set, connectivity goes out over managed VPC Lattice egress with no public endpoint anywhere, and the agent gets one search_tools meta-tool instead of 400 definitions - a problem the post puts at over 20,000 tokens per connection for a full tools/list.
The reported result is roughly 400 registered operations serving 10 internal agents across seven regions, in production for three months, with new API onboarding reduced to a spec annotation and a CI run instead of days of networking and security review. The caveat is that this is a vendor-customer post describing its own deployment, and the 'approximately six months of engineering effort saved' is MoEngage's estimate rather than an audited figure; the design also leans on AgentCore-specific pieces that do not port elsewhere unchanged.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.