CLOSE
aigocy@gmail.com

CUP

TWITTER (X) DRIBBBLE LINKEDIN

Configuration

COLORS
CUSTOM CURSOR
Skip to main content

espialtech

Menu
Start a Project

FlowForge

Home
/
Blog
espial
July 27, 2026
No Comments
FlowForge
Agentic Automation

Turn natural-language intent into multi-step, tool-using pipelines —
agents that re-plan around failure and escalate to a human when they are genuinely uncertain.

Project Details

The operations team ran on a patchwork of RPA bots, Zapier automations and manual handoffs. Every UI change broke something, nothing could reason about context, and there was no audit trail when a run went wrong. FlowForge replaced it with a planner-executor agent runtime. An LLM planner decomposes a plain-English process description into a typed DAG with explicit pre- and post-conditions per step — reviewable before it ever executes.

Executor agents call the existing APIs and tools over sandboxed REST and MCP connectors, with schema validation, retries and rate limits. A policy layer sits between them: confidence gates route low-certainty steps into a human approval inbox with one-click approve, edit or reject. Every tool call, prompt and decision is traced with OpenTelemetry, so any run can be replayed and audited.

DELIVERABLES
Agent runtime Tool orchestration Human-in-the-loop Full tracing
INDUSTRY
Enterprise Operations

Project Research

The interesting question was not whether an agent could execute a workflow — it was when it should refuse to. Too eager and it takes destructive actions on bad assumptions; too cautious and it escalates everything and saves nobody any time.

We calibrated the confidence gates against a labelled set of real historical runs, tuning the escalation threshold until the agent was aggressive where the evidence was strong and conservative exactly where a human would have hesitated.

Project Results

FlowForge removed 70% of manual operations work at a 99.3% run success rate — the self-healing behaviour is why: agents re-plan around a failed tool call instead of dead-stopping the run and waiting for Monday. New processes are built five times faster, because describing a process in English beats scripting it. And 100% of runs are auditable, which is what got the system past the compliance review.

HelpDeskAI
Support Agent
View Project

HelpDeskAI
Support Agent

DESCRIPTION
A support agent that resolves tickets end to end across chat, email and voice — 72% deflection, 38-second average resolution, and zero ungrounded replies.
DELIVERABLES
RAG architecture, Tool actions, Omnichannel, Escalation design
INDUSTRY
SaaS / Retail Support