Enterprise systems speak REST, SOAP, OData. Agents speak natural language. EAG is the layer in between — one canonical, plain-language tool catalog fronting every backend, governed on every call.
🧵
Build · Watch · Heal
Describe a system in plain language and EAG builds the connector itself — it probes the API, proposes the tools, tests them, and writes the manifest once a human approves. Then it watches every connection and repairs them when a backend shifts underneath you. The engine is Prava Loom, and it is governed by the same gateway it builds for: its own writes are attested in your audit trail as an agent, under the same ceiling as anything else.
🗂️
One Canonical Tool Catalog
A single, stable, natural-language contract agents route on. EAG maps each canonical tool to the right backend call. Descriptions are written for the model, so the agent's LLM matches them from any language. EAG itself never sees the user's language — only the resulting tool call — so execution and governance are language-independent by construction. Eval-verified: a Spanish request routes to the same tool, byte-identical, and returns the caller's own data.
🔐
Identity-Aware, Per User
Entra SSO in; OAuth On-Behalf-Of downstream where the backend supports it, so its own row-level security applies. Elsewhere EAG resolves the caller and filters to their rows — the identity on every call is the real user, never an anonymous “the app”.
🛡️
Governed by Tier
Read / write / privileged tiers with policy gates: confirm-before-write, privileged actions blocked or approval-gated. The agent can't do more than the caller's role allows.
✋
A Human Signs Off on Writes
A gated write parks instead of executing. The approver — a group, a named person, or the requester's own manager read live from your directory — decides in the console. Approval unlocks a single-use grant bound to the exact arguments, and the agent re-runs the identical call under the requester's own token: the write lands as them, never as EAG, which holds no credential that outlives a request. Change one argument and it is a new approval.
🔒
Signed Decision Receipts
Every outcome is signed with Ed25519 and written to a database you own — Dataverse, Azure SQL, Snowflake — carrying requester, approver, and a hash of the exact arguments. Alter a row and verification fails against the public key. Agents can certify their own decisions too, and the gateway stamps the requester from the verified token, so the “who” is never self-reported.
🪪
Know Which Agent Acted
The agent's identity comes from the token's azp claim — asserted by your identity provider, not claimed by the caller. A registered agent gets a ceiling intersected with its user's scopes, so a read-only research agent cannot write even when the person running it can. Registration can only ever narrow access, never widen it. Disable an agent and it stops for everyone, at once.
🧾
Audited End to End
Every call recorded — user, tool, system, outcome, latency. A ready-made compliance trail and seat-metering feed, exportable to your SIEM. The audit is the proof, not a vibe.
🔌
Any Backend, Even Legacy
REST, Workday RaaS, OData, SQL — modern APIs and legacy systems alike, added with a manifest and the console's Connect dialog (validate before you enable). Federate anything that already speaks MCP with zero driver code.
♻️
Migration Insurance
You license the catalog, not the backends. Swap ServiceNow → Jira or SAP ECC → S/4HANA one tool at a time, with a rollback toggle and an audit trail — and your agents never change.
🖥️
Admin Console Built In
One signed-in console runs the whole estate: flip systems on and off, connect backends with validate-before-enable, author tool-to-table mappings against live metadata, watch every call stream in, and manage user roles — no YAML required for day-two ops.
🔎
Identity Probe
Answer "what would this user see?" before they ask. Probe any scope set and get the exact tool list that identity would be served — allowed, gated, or hidden, with the policy reason — governance you can demonstrate, not assert.
🧪
Stage It, Then Scale It
No API yet? Model a target system in Dataverse, serve it live and governed through the same canonical catalog, and prove the agent experience on staged data. When you're ready, repoint the manifest to the real backend — and all agent traffic switches in seconds. Model → Stage → Prove → Repoint → Production.
🔑
One Image, Offline-Licensed
One container, one config-per-customer. A signed key unlocks your tier — verified offline, so it runs air-gapped with no phone-home. When it ships, every install will start with a trial — every feature unlocked.
