Live reference implementation · 240+ tool calls in production

An agent framework
built to survive contact
with the real world.

Most agent frameworks are optimized for the demo. KEVLAR is a set of patterns — deduplication, budget guards, silent-auth recovery, self-healing supervision — extracted from a voice assistant that has run continuously in production, not a weekend prototype.

Now shipped as a library → pip install kevlar-agent

240+
Orchestrated Tools
24/7
Continuous Runtime
5
Channels · One Core
0
Silent Failures Tolerated
Battle-Tested Patterns

Reliability isn't a feature.
It's what's left after things break.

Every pattern below exists because something failed first in a live system, with a live user on the other end. Here's what actually holds an always-on agent together.

01

Silent-Auth Recovery

OAuth flows that try to open a browser and wait for a human click will hang forever when called from an unattended background thread. Every credentialed call gets a silent mode that fails fast and reports instead of blocking.

Root cause: a scheduled task called a service helper that assumed a human was watching. It wasn't. The task hung silently for hours.
02

Duplicate-Call Suppression

A model can emit the same tool call twice in one turn — once from an eager first pass, once from a retry. A short time-windowed dedupe key (tool + args) absorbs the echo without touching the real side effect twice.

Root cause: two rapid duplicate calls to a "launch session" endpoint spun up two concurrent audio sessions on the same device.
03

Per-Feature Budget Guards

Every paid API call sits behind a monthly spend ceiling tracked independently per feature, not per account. A runaway loop in one capability can't drain the budget of every other capability sharing the same key.

Root cause: two features shared one provider key with no isolation; a spike in one silently starved the other of quota mid-conversation.
04

Confirmation as Code, Not Prompt

Irreversible actions require an explicit confirmed=true flag enforced in the dispatch layer — not a polite instruction in the system prompt. A model misreading intent once is not an acceptable failure mode for money or publishing.

Root cause: prompt-only guardrails are a suggestion the model can miss. Code-level gates aren't.
05

Full-Fidelity Audit Trail

Every tool invocation — args, result, error, latency, and calling channel — is appended to a structured log. Not for compliance theater: it's the dataset that reveals which capabilities are actually used, and which ones quietly fail.

Root cause: without a trail, "it worked when I tested it" is the only signal you have. That signal lies.
06

Self-Supervising Watchdog

A lightweight external process checks that every core service is alive on a fixed interval and restarts what isn't — independent of the services it watches, so one crash can't take down its own recovery mechanism.

Root cause: a process that only restarts itself can't restart itself after it's dead.
Now Installable

Every pattern above,
as a real package.

These six patterns aren't just documented on this page — they're shipped as kevlar-agent, a tested, zero-dependency Python library. Drop them into your own agent today.

  • dedupe, budget_guard, require_confirmation, silent, AuditLog, Watchdog
  • 24 tests, zero required dependencies
  • MIT licensed, published on PyPI
pip install kevlar-agent
$ pip install kevlar-agent
Successfully installed kevlar-agent-0.1.0

>>> from kevlar_agent import silent, SilentTimeout
>>> silent(refresh_oauth_token, timeout_seconds=15)
✓ returns fast, or raises SilentTimeout — never hangs
self_check.py — daily 05:30
Nightly Self-Audit

The agent checks its own pulse
before you have to.

A scheduled health pass exercises the real code paths — not just "did it import" — and only speaks up when something is actually wrong.

  • Background services verified alive via process inspection, not assumption
  • Every credentialed integration exercised in silent mode with a hard timeout
  • Tool schema integrity checked across every calling surface
  • Silent when healthy — a notification only fires when something needs attention
System Shape

One dispatch core.
Five very different front doors.

Voice, chat, and messaging channels all route through the same audited, budget-guarded dispatch layer — so a capability built once is trustworthy everywhere it's exposed, and every channel inherits the same safety rails automatically.

RealtimeVoice Session
ChatMobile Agent
MessagingTelegram
MessagingDiscord
TelephonyPhone / SMS
▼ ▼ ▼ ▼ ▼
Shared Core Dispatch · Dedupe · Budget Guard · Audit Log · Confirmation Gate
CapabilityCalendar / Email
CapabilityFinance / Ops
CapabilityContent / Media
CapabilityHome / Device
Capability240+ more…