Pre-launchEarly access October 2026 — reserve your spot →

Prove what your agent actually did.Stop the irreversible.

Verify every consequential action. Gate the irreversible.

One email when early access opens. No spam, unsubscribe anytime.

219 already on the waitlist

View on GitHub Follow the build
Claimed versus actual receipt, silent fail verdict, then an approval gate.
22+ integrations

One SDK. Every stack.

Works with every major agent framework.

  • LangGraph
  • LangChain
  • Agno
  • OpenAI Agents SDK
  • OpenAI SDK
  • Anthropic SDK
  • Claude Agent SDK
  • Google ADK
  • Microsoft Agent Framework
  • Semantic Kernel
  • AutoGen
  • OpenHands SDK
  • LlamaIndex
  • CrewAI
  • Pydantic AI
  • Mastra
  • Vercel AI SDK
  • Haystack
  • DSPy
  • AWS Strands
  • smolagents
  • OpenTelemetry
Built on OpenTelemetry — attach to the collector you already run
Control plane

See Agenko in action.

Full console preview — Observe, Control, Iterate, Setup. Click any nav item. Same plane you’ll run.

Join waitlist for early access ↗
console.agenko.app/traces
Agenko console
Observe
Control
Iterate
Setup
demo@agenko.app

Traces observations

Every run + verdict. Filter silent fails, open a trace for timeline and receipts.

Traces
128
24h
Silent fails
7
needs attention
Grounded
91%
hard facts
Spend
$18.4
today
WhenNameVerdictCost
41msupport-agentgrounded ✓$0.012
trace_9d44 · click a silent fail → receipts
What you measure

The evidence your dashboard misses.

Every consequential action leaves an exact payload, a gate decision, an observed result, or a clearly labeled estimate.

Silent-fail verdicts
claimed ≠ actual
hard fact on the receipt
Approval gates
before risk
irreversible actions pause
Root cause
graph walk
not an LLM guess
Audit trail
hash-chained
Article 12-ready export
One incident, end to end

A refund failure, made undeniable.

A single action record connects the trace, the decision, the execution result, and the next safe step.

  1. 01Trace

    issue_refund times out

    Agent claims success. Span shows 504 after 12s. No error surfaced to the user.

  2. 02Receipt

    claimed ≠ actual

    Verdict: silent failure. Diff on the receipt; Slack gets the alert.

  3. 03Gate

    Next refund held

    Irreversible retry pauses for human approval before money moves.

  4. 04Audit

    Hash-chained export

    Article 12 pack: who approved, what ran, what failed — one command.

Why Agenko

From scattered tooling to one action record.

Scattered tooling

Langfuse · LangSmith · Braintrust · Lemma · Agnost

  • Dashboards show latency, not whether the run was right
  • Silent failures sail straight through
  • No control in the execution path
  • High-volume, low-integrity span data
Agenko · the control plane

One action record · from intent to verified outcome

  • Per-run verdict with claimed vs. actual diff
  • Silent failures caught, pushed to Slack
  • Deterministic gate the model can't bypass
  • Low-volume, high-integrity, hash-chained
Where the boundary matters

Start with the actions you cannot quietly undo.

Agenko is most useful where an agent can sound confident, move fast, and still leave you with an expensive mess.

MONEY

Refunds & payments

Prove a refund settled before the agent tells the customer. Hold a retry when money is about to move again.

receipt + gate
IDENTITY

Deletes & data changes

Pause account deletion and sensitive writes until the right owner approves, with the decision attached to the trace.

approval required
PRODUCTION

Deploys & system writes

Show whether a release actually shipped, then stop the next irreversible write when the run diverges from its claim.

measured, not guessed
Why us

One control plane for the action lifecycle.

Traces, evals, gates, receipts, root cause, and audit evidence — connected around what the agent proposed and what the outside world received.

AxisDetectorsAgenko
ControlAlerts after the actionGates the action before it runs
EvidenceInferred failure scoresDeterministic receipt + trace
OwnershipClosed-source SaaS + OTEL + self-host
AuditSuggestions to investigateHash-chained export
Features

Prove the action. Control the risk.

The action record is the wedge. Receipts, gates, root cause, kill-switches, and audit evidence complete the loop.

Receipt

Did it actually work?

Per-run verdict with claimed vs. actual diff. Silent failures hit Slack.

Gate

Stop before the irreversible.

Deterministic block before deletes, payments, and deploys. Human approves.

Root cause

why(), not an LLM guess.

Deterministic graph walk to the step that diverged. No model call.

Kill-switch

Agents spend money while you sleep.

Budget, loop, and token caps. Hard stop per run. Free.

Tracing

OTEL-native, no lock-in.

Every span, tool call, token, and cost. Emits and ingests OTEL.

Audit

Evidence as a query.

SHA-256 hash-chained log. Article 12 export in one command.

12sinks & channels

Emit once. Route everywhere.

OTEL out to the collector you already run. Alerts and approvals land where your team already lives.

  • DatadogOTLP spans + metrics
  • HoneycombOTLP traces
  • GrafanaTempo / Loki via OTLP
  • Slacksilent-fail + gate alerts
  • DatadogOTLP spans + metrics
  • HoneycombOTLP traces
  • GrafanaTempo / Loki via OTLP
  • Slacksilent-fail + gate alerts
  • Discordwebhook notifications
  • Telegrambot alerts + approvals
  • WhatsAppbusiness API alerts
  • EmailSMTP approval requests
  • Discordwebhook notifications
  • Telegrambot alerts + approvals
  • WhatsAppbusiness API alerts
  • EmailSMTP approval requests
  • PagerDutyincident webhooks
  • HumanLayerapproval channel
  • OpsGeniealert integration
  • Microsoft Teamswebhook notifications
  • PagerDutyincident webhooks
  • HumanLayerapproval channel
  • OpsGeniealert integration
  • Microsoft Teamswebhook notifications
OTEL out to the collector you already run
How it works

One line in.
A receipt out.

Your first receipt in under 30 seconds. Add a gate, root cause, or a budget cap later - one decorator each.

agent.py
pip install agenko
import agenko

# receipt - prove the run did what it claimed
@agenko.check
def issue_refund(order, amount):
    return process(order, amount)

# gate - pause anything irreversible for a human
@agenko.approve(when="irreversible")
def delete_account(user_id):
    db.delete(user_id)
  1. 01

    Installpip install agenko

    One package, . Self-host in two containers or use managed Cloud.

  2. 02

    Get a receipt@agenko.check

    Add @agenko.check. Every call gets a verdict - claimed vs. actual. Silent failures ping Slack.

  3. 03

    Gate the risky calls@agenko.approve

    Wrap with @agenko.approve. Irreversible actions pause until a human approves - deterministically.

  4. 04

    Ask whyagenko.why()

    agenko.why(run_id) walks the trace to the step that diverged. A graph walk, not an LLM guess.

Before rollout

One control plane before the agent runs.

Four decisions. None of them need a new stack — just owners, a tool list, and where alerts land.

  1. 01Owners

    Who approves irreversible actions

    Finance for refunds. Platform for deploys. Privacy for deletes. A gate without an owner is a ticket that never closes.

  2. 02Guardrails

    Which tools pause

    Wrap the irreversible ones: payments, deletes, deploys, writes. Reads stay free. One decorator per action.

  3. 03Alerts

    Where silent fails go

    Slack for every silent fail. PagerDuty when they spike. Approvals wait in the inbox the owner already uses.

  4. 04Honesty

    Measured vs estimated

    Hard facts (tool called, args, status) are measured. Inferred scores stay labeled estimated. Never sell a heuristic as a fact.

Safety & compliance

Evidence built for production.

Tamper-evident records and export primitives for your governance program.

Tamper-evident audit

SHA-256 hash-chained log. Append-only, verifiable, exportable.

PII redaction, SDK-side

Sensitive data stripped inside the SDK before it leaves your infrastructure.

SSO, RBAC, self-hosting

Full on-prem deployment. Two containers, your cloud, your rules.

One-command evidence export

Signed evidence bundles for EU AI Act, NIST AI RMF, ISO 42001, SOC 2, GDPR, and emerging regulations worldwide.

Shubham Pitekar

Founder, Agenko

The scariest AI failures aren’t loud. They’re a 200 response that’s secretly wrong — and an irreversible action nobody paused.

I started Agenko after watching agents in production do the same two things: claim success when the consequential action never happened, and charge ahead on actions no one can undo. Teams could see everything and prove nothing.

So we’re building one open control plane around the action itself — the proposal, the decision, the execution, the evidence, the verdict. Self-hostable, built in public from day one.

shuence.com
FAQ

Questions, answered.

Everything you need to know before putting a gate in the path.

Product

  • What does Agenko actually do?

    Every consequential action gets two things. A receipt after it runs proves what the agent actually did versus what it claimed, with a verdict: grounded, ungrounded, or silent failure. A gate before it runs holds any irreversible action for human approval. Around both sit tracing, a free cost and loop kill-switch, deterministic root cause, and an audit-ready evidence export.

  • Is it available?

    Not yet. The SDK and platform are in active development, and early access opens October 2026. The spec and strategy are on GitHub, but nothing is released today. Join the waitlist to reserve your spot.

  • How is this different from Langfuse, LangSmith, or Braintrust?

    Those platforms are strong at traces, evaluations, and improving agent quality. Agenko adds a different primitive: an execution-boundary record that ties the exact tool call to its policy decision, observed outcome, and claimed-versus-actual verdict. It is open, OTEL-native, framework-agnostic, and designed to sit beside the tools you already use.

  • How does this compare to Lemma?

    Lemma focuses on auditing production traces, grouping recurring failures, and helping teams improve agents. Agenko focuses on the action boundary: capture the proposed tool call, hold or allow it before the side effect, then record what actually happened. The products can be complementary: diagnose the system, then prove and control the consequential action.

  • How does this compare to other detection tools?

    Detection and evaluation tools explain where quality is drifting. Guardrails and approval systems can restrict or pause behavior. Agenko connects those moments to one action record: what was proposed, what decision was enforced, what executed, and whether the result matched the claim.

Security

  • Will it be open source?

    Yes. The core is open source and self-hostable in two containers, the app plus Postgres, with OTLP out to Datadog, Honeycomb, or any backend. Managed Cloud is optional for teams that would rather not run infra. Self-host is free forever, and we monetize optional hosted support, never crippleware.

  • Are the verdicts real measurements or guesses?

    Both, and we label which is which. Hard facts, like whether a tool was called and with what arguments, are measured. Signals that infer rather than measure, like grounding scores, carry an estimated badge and a documented method. We never sell a heuristic as a measurement, and root cause is a deterministic graph walk over the trace, not an LLM guess.

DX

  • What is the developer experience goal?

    Dead simple. Call init() with no required arguments, let it auto-detect your frameworks, and your first receipt lands in under 30 seconds. The gate, root cause, and budget guards are one extra line each, only when you need them.

Prove the action. Control the risk.

The open execution boundary for consequential AI-agent actions. Apache 2.0.

One email when early access opens. No spam, unsubscribe anytime.

219 engineers already in the queue

View on GitHub→