Source expansion without chaos
The agent can scan a bounded set of known seeds and save candidate memory without turning everything into a live source.
EVAVO product system
Find better sources without letting automation run loose
Opportunity Agent is a private EVAVO product system: source intelligence, operator cockpit and review first automation layer in one. It is built to discover useful sources and opportunities, recover when source lists run dry and avoid uncontrolled scraping, generic outreach or AI noise.
Design thesis
The best agent is not the one that does the most. It is the one that knows what it found, why it matters, which method to try next and when to stop.
Mode
Product system
Default
Review first
Memory
Origin aware
Control
AI off by default
source intelligence console
Next recommendation
Rotate to sitemap scan
The latest bounded scan found links but no strong candidates. Try a safer alternate discovery path before increasing depth.
fallback: links without candidates
next: sitemap or query hints
guardrail: do not raise caps first
save path: candidate review only
Public link found
7 candidatesorigin=public_link_graph
Query resolved
4 candidatesorigin=query_hint
Sitemap found
8 candidatesorigin=sitemap
Expansion found
12 candidatesorigin=source_expansion
Problem
Opportunity discovery breaks when the source list runs dry
Most outbound and opportunity workflows depend on static lists, manual searching, shallow scraping or risky automation. EVAVO needed a system that could keep finding useful grants, tenders, supplier panels, digital service opportunities and partner openings without becoming noisy or unsafe.
Product idea
A source intelligence system that learns where to look next
Opportunity Agent starts from trusted sources, expands through bounded scans, reads public relationship signals, checks robots and sitemaps, supports query pattern discovery, preserves source origin and learns which paths create useful live sources.
Recovery model
When one method runs dry, the system rotates safely
The product does not just fail when sources are missing, stale, duplicated or low yield. It can fall back from trusted seeds to bounded scans, sitemap and robots discovery, public link graph checks, query hint recovery, candidate review and source health judgement before more budget is spent.
Control model
Everything risky stays behind review gates
AI calls, email sending, source promotion, source health actions and candidate saves are separated from discovery. The browser never receives the Worker admin token, and the safest mode keeps AI and sending off while source intelligence continues to improve.
Discovery architecture
The source intelligence flywheel
The system separates source discovery from source promotion. That makes it safer, easier to audit and easier to recover when one discovery method runs dry
01
Trusted source seeds
02
Bounded source expansion
03
Worker fallback guidance
04
Public link graph discovery
05
Robots and sitemap discovery
06
Query pattern source hints
07
Human reviewed URL resolving
08
Scored candidate memory
09
Candidate review brief
10
Source health judgement
11
Confirmed source promotion
12
Origin metrics and strategy learning
UX layers
The experience is split by audience
A strong AI product is not one screen. It has a public story, a private operator cockpit and a backend intelligence layer that can be trusted
01
Public product narrative
Clear enough for clients to understand the offer without exposing private operational detail.
02
Private Ops cockpit
A dense but readable control centre for setup, discovery strategies, source health, review, learning and audit trails.
03
Worker intelligence layer
D1 memory, strategy scores, origin preservation, run records, source candidates, public link candidates, fallback guidance and guarded admin endpoints.
Capability design
Built like a product, not a demo
The system combines backend controls, data memory, recovery paths, review surfaces, safety gates and public product storytelling
Fallback source recovery
When a scan finds nothing, hits failures, sees thin pages or rediscovers duplicates, the system can recommend the next safe method instead of blindly increasing depth.
Public link graph discovery
It can inspect ordinary public links from known pages for partner, client, supplier, funding, portfolio and opportunity style source candidates.
Sitemap aware discovery
It can use robots and sitemap files to find procurement, grant, supplier and funding pages that are not obvious from a portal homepage.
Query hints with human resolving
It can generate high intent AU and NZ query patterns, then let an operator paste useful result URLs back into the system for scoring and dedupe.
Candidate and health judgement
Candidate batches and source health states are interpreted before promotion, so the operator knows when to promote, inspect, tighten, pause or hold.
Origin aware source memory
Saved sources preserve where they came from: query hint, public link graph, sitemap, bounded expansion, preview or manual entry.
Adaptive recommendations
The cockpit shows whether to increase, maintain, tighten, seed or cool down a discovery path based on yield, health, duplicates, failures and review outcomes.
Audit first operations
Runs, source results, candidate rejections, score diagnostics, reviews, learning passes and route contracts are visible from the Ops cockpit.
Next
A reusable intelligence layer for EVAVO style products
This is the kind of work EVAVO should be known for: polished systems that combine brand, UX, operations, automation, source recovery and product safety. Not just websites. Not just AI prompts. Work done properly.