Real-time trace, replay, and self-heal production LLM pipelines. Detect prompt drifts and infinite agent tool loops with sub-millisecond latency.
Powering 5M+ daily agent traces across leading AI startups.
Trusted by engineering teams at
Agent trace tree
tool.refund_lookuptool
Loop detected on call 4 (same args x4). Circuit breaker tripped and rerouted to fallback model.
2,210 tokens
Latency
now 38ms
38ms avg
Token cost
$0.021 / run
62% under budget after fallback routing
$ npm i @pulsestack/sdk
+ @pulsestack/sdk@3.0.2 · tracing enabled in 1 line
Failure modes
Loops get cut, every decision can be replayed, sensitive data is masked before it leaves your network, and every request lands on the cheapest model that still passes.
Circuit breakers
Detects repeated tool calls with identical arguments and trips a breaker in-flight. The span reroutes to a fallback instead of burning retries and tokens.
pulse.breaker({
maxRepeats: 3,
onLoop: (span) => span.terminate({
fallback: "small-model",
}),
});
// tool.search_orders(48213) x4
// -> loop detected, run terminated in 38msReplay
Step backwards through every agent decision. Prompts, tool responses and seeds are recorded, so any run replays exactly.
Step 4 of 6 · tool.refund_lookup
6,820 → 2,592 tokens / run
Compliance
Sensitive fields are masked at the edge, before traces are stored or sent to any model provider.
Routing
Smart fallback routing sends each request to the cheapest model that passes your evals, and fails over automatically when a provider degrades.
Engine
PulseStack sits beside your agent runtime, not inside your prompts. Every step is traced at sub-millisecond overhead, checked against rules you can read, and repaired by policies you version like code.
Built OpenTelemetry-native, so existing exporters keep working. First-class adapters for LangChain, LlamaIndex, and the OpenAI Agents SDK.
Stage 01 · Capture
One call instruments your agent graph. Prompts, tool calls, arguments, token counts, and latency are emitted as OpenTelemetry spans with under 1 ms of overhead per step.
import { pulse } from "@pulsestack/sdk";
const agent = pulse.wrap(supportAgent, {
project: "support-prod",
otel: true,
});Stage 02 · Detect
Detectors run inline on the span stream. Prompt drift is scored against your baseline, and repeated tool calls with identical arguments trip a loop rule before they burn budget.
detect "tool_loop" {
when same_tool_args >= 4 within 30s
scope tool.*
action trip_circuit_breaker
}Stage 03 · Self-heal
Healing policies are plain YAML checked into your repo. A tripped breaker can fall back to a cheaper model, retry with a corrected prompt, or mask PII at the edge before the response leaves your network.
healing_policy:
on: [tool_loop, prompt_drift]
fallback_model: gpt-4.1-mini
max_retries: 2
pii_mask: [email, card_number]Benchmarks
The numbers production teams hold us to, on the same workload every release.
From span receipt to on-call notification.
Added per traced LLM call, measured in-process.
Trailing 12-month ingest availability.
Across production agent fleets.
PulseStack
0.4s
Circuit breaker trips on the fourth identical call.
Log-grep tooling
47 min
Saved queries over raw logs, run on a schedule.
Manual review
3+ hours
Engineer reads transcripts after a customer reports it.
Bars are drawn to scale against the manual baseline. The PulseStack bar is widened to a visible minimum.
Methodology: all figures measured on a 10k-trace synthetic workload, 8 vCPU. Latency is p50 unless noted. Uptime is trailing 12 months of ingest availability.
Savings model
Move the sliders to match your traffic. Circuit breakers recover about 85% of wasted spend, and fallback routing saves roughly 30% on what remains.
Your workload
Assumes circuit breakers recover 85% of wasted spend and fallback routing saves 30% on remaining spend. Scale plan at $79/mo is deducted from net savings.
Net monthly savings
$7,349
After the Scale plan at $79/mo
Founder reviews
“We were paging on-call three times a week for agents looping on the same refund lookup. PulseStack trips the circuit breaker on the fourth identical call and reroutes to a fallback model before a customer ever sees it. Agent incidents dropped by more than two-thirds in the first month, and our on-call rotation finally sleeps.”
Priya Raman
CTO, Kestrel Labs
-71% agent incidents
“Time-travel replay turned a two-day hunt for a prompt drift into a twenty-minute read. We step back through each planner decision and see exactly which system message changed.”
MTTR 4h → 6min
Tomasz Nowak
Head of Platform, Lattice Harbor
“Edge PII masking was the gate for our healthcare pilot. Emails and card numbers never leave our VPC, and the compliance review took a week instead of a quarter.”
$18k/mo saved
Dana Whitfield
Founder, Quorra Systems
Photos by Pexels contributors
Plans
Every tier runs the same tracing engine. You pay for volume, retention and the controls your production agents need.
Developer
For side projects and local agent experiments.
Scale
For production agents shipping to real users.
Enterprise
For regulated teams with strict data-residency needs.
All plans include a 14-day Scale trial. No credit card required.
Developer FAQ
Latency budgets, data residency, framework coverage, and how the breaker makes its decisions. Specifics, not marketing.
Still stuck? Talk to an engineerInstall the SDK in one line and get traces, replays, and circuit breakers on your first production run.
14-day trial. No credit card required.
$ npm install @pulsestack/sdk