A practitioner's guide from PrometheanLink / AgentOceans — the people who run this every day.
We built our agentic system over a year of running it for real — invoicing clients, shipping software,
monitoring infrastructure, and handing off across sessions. What follows is how it actually works.
A chatbot answers a question. An agent has to carry a long job across hours, context
limits, and interruptions without dropping the thread, lying about what's done, or taking an action it can't
take back. The difference between the two isn't the model — it's the harness around it.
Here's the seven-layer harness we run, in plain terms. Each layer is a place long-running agents fail — and
the discipline that keeps them from failing there.
Layer 1 — ContextRead before you trust.
Load the project rules and the last handoff — then inspect the files it references before believing its
status. A handoff that says "done" isn't done until the work is on disk. Resolve conflicting instructions first.
Layer 2 — ProcedureEvery step leaves a mark.
Read sources → build the artifact → check it against acceptance criteria → fix failures and save the report →
record the next action. If a step has no observable output, it didn't happen.
Layer 3 — ToolsConfirm the rail before the long run.
Prove one source is reachable before committing to a long task. Record retrieval failures. Retry only after
changing something. If a tool's down, name the blocker and save the work that can still be finished.
Layer 4 — PermissionsProse is not a boundary.
Keep edits in scope with real controls, not good intentions. Ask before publishing, sending, or spending.
Before retrying an external write, check whether the first one already landed. Keep a recoverable version before
a risky edit.
Layer 5 — ReviewA separate pass, with evidence.
Grade the work against each requirement in a table — and leave missing evidence visible, not hidden:
Layer 6 — EffortSpend thinking where it's hard.
Use the configured model for the main task; reserve deeper reasoning for ambiguous or costly decisions. Don't
pretend a prompt changes runtime effort — measure, don't assume.
Layer 7 — CompletionDefine "done" before you start.
Write the completion checklist up front, each item pointing to observable evidence: the output opens, the checks
ran against the saved version, failures are fixed or reported, the handoff matches the files. Blocked? Return a
partial status with the exact remaining work — never a silent stop.
The loop that ties it together: inspect saved state → take one bounded step → build → check → keep good
work, undo only failures → checkpoint → continue to completion, a blocker, or the run limit. Never repeat a
failed approach without a new reason. After each meaningful stage, write the handoff — the task, the exact
output paths, what's done, the decisions and why, the open issues, and one concrete next action.
Why this matters
Most "AI agent" failures aren't intelligence failures — they're harness failures. The agent forgot what it
already did, trusted a stale note, took an action it couldn't undo, or declared victory without proof. The seven
layers are just the places to stand guard. Get them right and an agent can carry work across a context limit the
way a relay runner carries a baton — nothing passes but the thread, and the next runner picks up mid-stride.
AgentOceans · Agentic Systems Harness — setup & consulting
We build these for a living.
AgentOceans runs a multi-agent harness that invoices clients, ships software, monitors infrastructure, and
hands off across sessions — every day. If your company wants an agentic system that actually finishes long work
instead of demoing it, we'll help you stand up your own harness, wired to your tools and your guardrails.
✓
Thanks — we'll reach out about your harness setup.
Your request is in. Expect a reply by email from the PrometheanLink team at the address you gave us.
AgentOceans — a PrometheanLink system · PrometheanLink LLC · agentoceans.com