FOR HUMANS YOU ARE HERE
FOR AGENTS READ discovery.md
OBSERVE · EXTRACT · REPLAY
DISCOVER · MIRROR · CAPTURE
2026

DISCOVERY.
MIRROR ANY SYSTEM

any system. any process. any engine.

An agent that walks into any existing system, observes how it really behaves, captures the rules and workflows that govern it, and mirrors the whole thing back inside ERP.AI as a live Proto. You don't migrate. You discover.

Run a Discovery How it works
Live Discovery [ DISC-3309 ]
1,284 Entities mirrored
317 Rules captured
42 Workflows observed
Phase Discover Observe Capture Replay
Snapshot [ ACTIVE ]
  • Source Any System
  • Method Observe-Extract
  • Output Live Proto
  • Status Live
Agent Class DISC-X1
DISCOVERY • OBSERVE • EXTRACT • MIRROR • CAPTURE • DISCOVERY • ANY SYSTEM • ANY PROCESS • ANY ENGINE • REPLAY IN PROTO • DISCOVERY • OBSERVE • EXTRACT • MIRROR • CAPTURE • DISCOVERY • ANY SYSTEM • ANY PROCESS • ANY ENGINE • REPLAY IN PROTO •
01
DISCOVER
Walk the system, map surface
02
OBSERVE
Watch real workflows in motion
03
CAPTURE
Extract rules, decisions, paths
04
REPLAY
Mirror as a live Proto agent
Cycle Pillars Sources Architecture Evaluation Cases Replay
01 Three Pillars

Discover. Mirror.
Capture .

The Discovery agent doesn't ask for documentation. It reads the system the way a senior operator would: by following real work through it, watching what gets approved, what gets rejected, what gets escalated, and which fields actually matter. Three modes, one mission — hand Proto a faithful clone of how the thing really runs.

Pillar 01

Discover .

Crawl the source system end-to-end. Surface every screen, schema, queue, and integration point. Build a topology of what exists before deciding what matters.

Status: Mapping
Pillar 02

Mirror .

Replicate the structure inside ERP.AI: the same entities, the same states, the same approval chains. The mirror starts hollow and fills as Discovery learns.

Status: Replicating
Pillar 03

Capture .

Pull the implicit rules: who approves what over which threshold, which exceptions get auto-routed, the conditions humans actually check. Encode them as Proto policies.

Status: Encoding
02 Source Surface

Any system. Any process.
Any engine .

No connector library. No vendor list. No exceptions.

Any System

Source-of-truth, system-of-record

From a thirty-year-old mainframe to last quarter's internal tool. If a person can use it, the Discovery agent can read it. UI, API, database, file drop, whatever surface the system actually exposes.

Any Process

How work moves through the org

Sequential flows. Branching approvals. Out-of-band Slack threads that finalize decisions. The agent watches end-to-end runs, not just the happy path, and captures the variants that humans treat as routine.

Any Engine

The rules layer underneath

BPMN, rule engines, hard-coded conditions, spreadsheet logic, tribal knowledge. Discovery extracts the actual decision logic in use, including the rules that aren't written down anywhere.

Trace: Every observation, extracted rule, and mirrored entity is logged and replayable
03 Architecture

Three subsystems
one coordinator .

Discovery is a Proto-class agent. It inherits Proto's primitives, self-generating topology, runtime tool synthesis, persistent knowledge graph, multi-agent coordination, and adds three Discovery-specific subsystems. The Coordinator owns the trace and the topology; the subsystems do the work.

Subsystem 01

Surface Adapter

Reads whichever surface the source exposes: UI, REST, GraphQL, read-only DB, file drop, queue tail, mailbox. Synthesizes adapter code at runtime if no native client exists.

Output: Normalized event stream
Subsystem 02

Mirror Allocator

Maintains the shape-capture graph: entity types, fields, states, relationships. Allocates ERP.AI-side equivalents and tracks the source-to-mirror mapping.

Output: Live entity graph
Subsystem 03

Rule Extractor

Watches state transitions and human decisions. Hypothesizes conditions, tests them against held-out events, promotes confirmed rules to candidate Proto policies.

Output: Candidate policy set
Layer

Coordinator

Inherited from Proto. Owns the topology, schedules sub-agents, merges their outputs, writes the replayable trace. Decides when to spawn, merge, or dissolve.

Output: Trace · DISC-id
04 Evaluation

Did the mirror really mirror ?

Four metrics. One question.

Illustrative, pre-launch: targets shown are v1.0 defaults; v1.1 will replace with measured production data

Coverage

≥ 90%
of high-volume scopes mirrored

Mirrored workflow scopes / total inventoried scopes, weighted by event volume.

Fidelity

≥ 0.92
F1 vs. human decisions

Captured-policy decision agreement against human decisions on held-out events.

Replay accuracy

≥ 95%
match on fresh sample

Proto's end-to-end execution vs. source-system execution on the same fresh inputs.

Drift

± 5 pts
rolling 14-day F1

Decay rate of policy agreement over time. Drift triggers re-capture, not silent decay.

Higher-stakes scopes (anything touching payments) ship with stricter defaults. See whitepaper Chapter 7 for the full framework.

05 Case sketches

What Discovery actually
caught .

Three anonymized walk-throughs from the restricted preview. Vendor names, customer identities, and absolute volumes are withheld; relative figures and substantive findings are preserved. Full case detail is in the whitepaper.

Case A

Major legacy ERP · AP module

96.4%
Replay accuracy

Finding: Email-mailbox surface was load-bearing. The 'official' workflow had one approval step; the actual workflow had a parallel email-based escalation path used for time-sensitive cases. No documentation captured it. Discovery did.

Case B

Tier-1 SaaS CRM · lead routing

14
Operator-level inconsistencies surfaced

Finding: Same lead profile was being routed to different reps depending on which manager was paying attention. Suspected for years, never quantified. Discovery quantified it; the remediation was a sales-ops decision, not a technical one.

Case C

Spreadsheet approval pipeline

97.0%
Replay accuracy

Finding: Cell J42 had a comment thread from three years earlier explaining a rule change the team had forgotten. Discovery surfaced the comment as evidence and asked whether the rule still applied. It didn't; the captured policy was corrected before promotion.

Pre-launch projections drawn from preview-cohort engagements. Full case detail in whitepaper Chapter 8.

06 Whitepaper

The full case
for Discovery .

Architecture in operational depth. The four-phase loop with budgets and exit criteria. The full evaluation framework with synthetic-benchmark targets and failure modes. Three anonymized case sketches with the numbers. Handoff contract. Safety envelope. Glossary. Twenty-plus pages of substance, not a brochure.

Document DISC-WP-001
Version 1.0
Released April 2026
Format PDF · Letter
Download whitepaper · PDF Open in browser

Quantitative figures marked illustrative are pre-launch projections; v1.1 will replace with measured production data.

Contents [ 10 chapters ]
  1. 01 Introduction: the migration problem
  2. 02 The Discovery thesis
  3. 03 Architecture overview
  4. 04 The four-phase loop, in detail
  5. 05 Source surface taxonomy
  6. 06 Rule capture methodology
  7. 07 Evaluation framework (illustrative)
  8. 08 Anonymized case sketches
  9. 09 Handoff to Proto
  10. 10 Guardrails, safety, and human gates
+ Glossary & references
07 Replay

Captured
Live Proto .

Discovery hands its capture to Proto. The mirrored entities become Proto's working data. The captured rules become Proto's policies. The observed workflows become Proto's topology. What ran in the legacy system runs as a Proto agent: same behavior, replayable trace, human override gates.

ERP.AI Discovery / Hands off to Proto
Capture Proto [ HANDOFF ]
Entities
Mirrored
Rules
Encoded
Workflows
Topology
Output
Live Proto
Discovery completes when Proto can run the workflow end-to-end without human help
08 Start a Discovery

Point it at a system it does the rest

One API call kicks off a Discovery run. Provide an access surface (URL, credentials, or read-only DB connection), the slice of the business you want mirrored, and a budget. The agent reports progress as it discovers, mirrors, and captures.

start-discovery.sh 
# Mirror an existing system into a live Proto
curl -X POST https://api.erpai.studio/v1/discovery/start \
  -H "Authorization: Bearer $ERPAI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "source": {
      "kind": "any",
      "access": "https://legacy.example.com",
      "credentials_ref": "vault://disc/legacy"
    },
    "scope": "accounts-payable",
    "depth": "workflows+rules",
    "budget_tokens": 100000,
    "human_approval": ["mirror_create", "rule_promote"],
    "target": { "proto": "auto-create" }
  }'

# Response
# {
#   "discovery_id": "DISC-3309",
#   "status": "discovering",
#   "mirrored_entities": 0,
#   "captured_rules": 0,
#   "proto_id": "PRT-pending",
#   "trace_url": "https://app.erpai.studio/trace/DISC-3309"
# }

* Read-only by default. Write-back to the source system requires an explicit human-approved policy.

Discovery Agent Active

Mirror Any System Into
A Live Proto.

Point Discovery at a system. It hands Proto a faithful clone: workflows, rules, and all.

Run a Discovery Back to Proto

Any system · Any process · Any engine · Read-only by default · Full capture trace