AGNT5 is the durable runtime and observability layer for AI workflows and agents.

We help teams move from brittle PoCs to production systems by handling the hard parts: durability, retries, state, traces, and evals — so your workflows recover instead of fail.

• Durable AI workflows
  Long-running, multi-step flows that survive crashes, restarts, and deploys.

• Agents with tools & human handoffs
  Orchestrate LLM calls, tools, APIs, and human approvals as one workflow.

• Context & RAG pipelines
  Ingestion, enrichment, chunking, retrieval, and re-ranking with retries and state built-in.

• Evals & experiments in production
  Compare prompts, models, and configs using real traffic, not just notebook data.

AGNT5 gives you production primitives most teams end up rebuilding:

• Durable runtime – turn regular async functions into workflows.
• State & idempotency – persist state between steps; safely retry work.
• Retries & schedules – declarative retry policies and cron-like schedules.
• Signals & human approvals – wait for external events or approvals as first-class steps.
• Observability – traces, logs, and metrics for every run.
• Eval hooks – attach evals to workflows to catch regressions early.

from agnt5 import durable, Context

@durable("triage_ticket", retries=5)
async def triage(ctx: Context, ticket: dict):
    payload = {"role": "user", "content": ticket["text"]}

    # Step 1: draft a reply with an LLM
    draft = await ctx.step("draft", lambda: ctx.llm.chat(messages=[payload]))

    # Step 2: wait for human approval (can pause for hours or days)
    decision = await ctx.signal.wait("agent_approval", timeout="24h")
    if decision != "approve":
        return "held"

    # Step 3: send the email
    await ctx.step("send", lambda: ctx.email.send(
        to=ticket["user"],
        body=draft.text,
    ))

    return "sent"