HYDRADB HACKATHON · TRACK 2 · LIVE CONTEXT FOR AGENTS

Your agent loads 329 tools……to send one email.

›Email Acme Corp's March invoice to their finance contact
THE OTHER 323 TOOLS NEVER ENTER THE CONTEXT
CONTEXT 27,000 tokens · 329 tools loaded

A tool-routing graph for agents, built on HydraDB — and it learns from every outcome.

Open the product → See the evidence ↓
SCROLL · THE EVIDENCE
01 · THE RACE

Same task. Three ways to load tools.

A vector store guesses similar tools. A chain is not a similarity problem.

PRELOAD-ALLthe default today
27,000 tok / task
chain “contained” — at full price
TF-IDF TOP-Kvector-style retrieval
823 tok / task
full chain only 8/20
WAYPOINT GRAPHtraversal on HydraDB
~254 tok / task
full chain 20/20 · exact
20 benchmark tasks · known chains 106× fewer tokens than preload bars on a log scale — the gap is bigger than it looks
02 · IT LEARNS

The graph rewrites itself from outcomes.

Every outcome rewrites the graph's weights. Failed chains lose their edges.

TOP-CHAIN ACCURACY · 4 PASSES OVER THE SAME 20 TASKS
0 .25 .50 .75 1.0 pass 1 pass 2 pass 3 pass 4 0.40 0.35 0.85 0.90 0.95 → 0.95
■ v2 · length-first ranking + outcome feedback - - v1 · weight-first — first attempt, failed, kept honest
FIRST ATTEMPT · V1
0.40 → 0.35

Weight-first ranking got captured by long detour chains. We published the failure — then fixed the rule, not the numbers.

WITH FEEDBACK · V2
0.85 → 0.95

Success strengthens a chain's edges (+1); failure weakens them (−1). One run, no tuning — every move on the curve is explainable.

03 · IT HOLDS UP

50 agents at once. It broke. Then it didn't.

The load test broke it. The fix is in the architecture, not the excuses.

BEFORE · NAIVE PER-CALL QUERIES
126
ERRORS / 150 CALLS
0.61
CALLS / SEC
→
AFTER · SNAPSHOT CACHE + LIVE TRAVERSAL
0
ERRORS / 150 CALLS
33.2
CALLS / SEC
50 concurrent agents · 3 rounds each 54× the throughput after the fix same box, same protocol — before and after both published
04 · YOUR TOOLS, YOUR GRAPH

Point Waypoint at your MCP servers. It builds the graph itself.

No hand-built catalog. Ingestion reads your tool schemas and wires the graph — proven live on the official reference servers.

OFFICIAL MCP REFERENCE SERVER · INGESTED LIVE
filesystem 14 tools
read_file→write_file a copy task, routed live · 230 tok
OFFICIAL MCP REFERENCE SERVER · INGESTED LIVE
memory 9 tools
search_nodes knowledge-graph lookup, routed live · 122 tok
✕ STOPPED and a read-only role is stopped at write_file — the graph knows who may call what.
05 · INSTALL

Ships as an MCP server.

Your agent connects in one config block — and holds three meta-tools instead of your whole catalog.

MCP CLIENT CONFIG
{
  "mcpServers": {
    "waypoint": {
      "command": "python3",
      "args": ["app/mcp_server.py"]
    }
  }
}
find_tools · the exact chain for the task resolve_entity · entities already known to the graph report_outcome · the write-back that makes it learn