updown move enter open esc close matheesan.ca / index

    CASE 00 · agentic tooling

    The agentic workbench

    At Replicant: one engine for the whole org — data silos, plugins, MCPs and gated LLM work, built solo.

    solo buildorg-wideovernight loops
    agentic toolingMCPpluginsDeepSeekClaudeJIRAgates
    System map of the agentic workbench: data silos connect through a plugin bus to the engine, DeepSeek and Claude adapters, gates, and gated plan, scope, build, JIRA and review outputs.
    workbench / system map · not to scale

    problem

    Engineers were working across disconnected data silos and separate streams of work. AI was useful per task, but there was no shared operating layer that could plan, scope, build and organize the work across those systems — and nothing could safely run overnight.

    approach

    Designed and built a single agentic engine from scratch: a plugin layer and MCP connections for multiple data silos and workstreams, a common work graph, and LLM adapters for DeepSeek and Claude. Wrapped it in my own harnesses and a gated loop: plan, scope, build, organize JIRA, verify, request approval, then continue. It could run overnight without turning a plausible output into a silent change.

    the hard part

    Unattended agents are a trust problem, not a prompt problem. Every action needed permissions, durable state, idempotency, an audit trail and a gate that could stop it. The engine had to stay legible when it was moving across systems, and it had to fail closed when a dependency or a model response was wrong.

    result

    Gave engineers across the org one engine to organize multiple data silos and streams of work. Work that used to be scattered across tools — planning, scoping, building and JIRA organization — could be handed to an overnight loop and reviewed in the morning behind gates. It leveled up how the org used AI, not just how one team shipped.

    what I’d do differently

    I would productize the gate and audit surface first. The engine was only as useful as a person’s ability to see what it did, trust it and stop it.