The 60-second failure mode

The fastest way to see why a naive tool-using agent loop breaks down — and what contextweaver does about it — in one command, with no API keys and no network.

contextweaver demo --scenario killer

(Also available as python -m contextweaver demo --scenario killer.)

The scenario

An internal ops agent with 100 tools and a running conversation. The user asks:

"Find unpaid invoices, check the account notes, and draft a reminder."

A naive loop pays for three things at once:

1. The tool catalog — all 100 tool descriptions injected into the route prompt.
2. The conversation history — every prior turn included raw.
3. A huge tool result — the invoice/account dump pasted straight back into the answer prompt.

contextweaver narrows each one:

	Naive	contextweaver	Reduction
Tools in the route prompt	all 100 descriptions (6,326 chars)	5 ChoiceCards (491 chars)	92.2%
The huge tool result	raw (14,430 chars)	firewalled summary (60 chars)	99.6%
The full answer prompt	everything raw (21,332 chars)	compiled (814 chars)	96.2%

Sizes are reported in characters (deterministic everywhere). The demo also prints a token estimate using the active tokeniser.

What you are seeing

• Route narrows the catalog. Router.route(query) turns 100 tools into a 5-card shortlist — the Tool Router at work. The model never sees 100 schemas.
• The firewall externalises the big result. The ~14 KB invoice dump is stored out-of-band as an artifact and replaced with a short summary — the Context Firewall. The raw bytes stay addressable.
• The answer prompt is compiled, not concatenated. A budget-aware build keeps the relevant history plus the summary, instead of dumping everything.

Where to go next

• The catalog showcase architecture is the same story as a runnable, inspectable script with BuildStats.
• The Showcase walks the other demo --scenario flows (large-catalog, huge-tool-output, mcp-gateway-full).
• The Quickstart shows the direct-API version you would embed in your own agent loop.