01Platform · Human-in-the-loop · MCP

Zasti Agentic Automation Platform

A shared agent harness (LangGraph through MCP) with four automation pipelines built on it — sales, resource tracking, marketing, and internal software engineering — adopted company-wide including leadership, with a human approval gate before any consequential action.

Zasti Inc · Design and implementation

4
verticals automated: 1
shared agent harness: 10
person company, adopted end to end: 0
consequential actions without human approval

In short

The work was too judgment-dependent for rule-based RPA and too repetitive to justify expert attention on every instance, so the platform was built for autonomy in execution and human authority in approval.
One reusable harness rather than four bespoke bots: the common layer handles task decomposition, tool use, and state, so each vertical inherits orchestration instead of reimplementing it.
Agents first fell into non-convergent loops — re-planning and re-invoking tools without ever reaching a terminal state. The fix belonged in task scoping, not model quality: work was restructured as an explicit task-dependency map, with each agent holding exactly one task at a time instead of reasoning about the whole goal.
The second failure mode was hallucinated output, addressed with layered validation gates that checked agent output before it could progress to a human reviewer.
Adopted across the entire 10-person company — every team plus leadership — absorbing work that would otherwise have required dedicated headcount in each of the four verticals.

Situation

Across Zasti, a large share of skilled people's time went to repetitive, high-volume process work — sales pipeline handling, resource tracking, marketing production, and internal software-development chores. The work was too judgment-dependent for classic rule-based RPA, but too repetitive to justify expert attention on every instance.

The mandate

Automate these workflows without removing human judgment from decisions that carry real consequences: autonomy in execution, human authority in approval — across four different domains with one reusable harness rather than four bespoke bots.

What I built

A shared agent harness, with LangGraph reached through MCP — a combination rather than a single framework, selected per pipeline.
Four pipelines on top of it: sales automation, resource tracking, marketing, and internal software-engineering agents for the dev team.
A common layer handling task decomposition, tool use, and state, so each vertical inherited orchestration instead of reimplementing it.
Deliberate human-in-the-loop design throughout: agents prepared and staged work, a person approved before anything irreversible executed.

The hard part — non-convergent loops

Agents fell into infinite loops: re-planning, re-invoking tools, never converging on a terminal state. Reading current practitioner writing and interviewing researchers and engineers working on agent reliability both pointed the same way — the failure was task scoping, not model quality. I restructured the work as an explicit dependency map and forced each agent to hold exactly one task at a time rather than reasoning about the whole goal at once, which removed the re-planning cycles.

The second failure mode — hallucinated output

Agents produced confident but wrong output. Layered validation gates checked agent output before it could progress, so hallucinations were caught before they reached a human reviewer rather than after.

Outcome

Adopted across the entire 10-person company — every team plus leadership used the pipelines. At that size there is nowhere to hide an unused tool.
The platform absorbed work that would otherwise have required dedicated headcount in each of four verticals. The claim is avoided dedicated hires, deliberately not a dollar figure.
Loop failures were eliminated via single-task decomposition; hallucinated outputs were caught by validation gates before reaching a human.

Why it generalizes

One harness, four consumers. It is platform thinking rather than one-off scripting: build reusable internal infrastructure that frees domain experts from repetitive work while the expert keeps judgment.