Paniolo Wiki
Keep your knowledge correct while the work moves underneath it.
Explore Wiki
Your knowledge. Your rules.
Keep knowledge useful
as the work moves.
Your knowledge and agent setup belong beside your work: versioned, reviewable, and ready to change with it. Start with the Windows or Linux x64 pilot.
Then preview setup without changing files
paniolo init --dry-runPaniolo Harness
Audit and evolve the rig your agents work in.
Explore HarnessInformed by published research.
Agentic Harness Engineering (AHE), by Lin and colleagues, studies how coding-agent harnesses can improve through a cycle of observation, changes, and evaluation while keeping the base model fixed.
In the researchers' experiments, ten iterations increased first-attempt task success on Terminal-Bench 2 from 69.7% to 77.0%. The evolved harness also transferred to SWE-bench-verified, using 12% fewer tokens than the starting harness. These are results from the paper's experiments.
The lesson for Paniolo is practical: the knowledge, tools, and operating environment around an agent deserve engineering discipline. We use research like this to inform how we organize agentic knowledge, strengthen harnesses, and test whether changes improve the work. The bibliography — 288 digested papers, including what we do not claim — is on the paper index, with product-by-product analysis on the research pages.
Terminal-Bench 2
SWE-bench-verified
AHE · 2026
Cross-model Transfer
Tools · Middleware · Memory
From the Founders
Follow on Substack
Harnessed
Obsidian is the human half of an LLM wiki
Karpathy's gist makes the model the programmer and the wiki the codebase. This essay is about the other seat: Obsidian as the place a person reads the same markdown the agent writes.
Herding agents
One terminal for Claude, Codex, Devin, and the rest. The hard part of several coding agents is knowing which pane is working, which is blocked, and which finished twenty minutes ago.
A Review of Agentic Harness Engineering
What changes when the model stays fixed but its tools, middleware, and memory improve? Our review explores Agentic Harness Engineering and what its findings mean for building a better environment around coding agents.
Agent-First Is the New Mobile-First
We put WebMCP on our own contact form, tested the workflow, and captured the lessons in a reusable skill. See how agentic knowledge and harness engineering connect.
Choose where to begin.
Build the intelligence layer
your agents need.
We help teams organize agentic knowledge, strengthen coding harnesses, and expose reliable tools for customer-facing workflows. Tell us what your agents need to know, what they need to do, and where they get stuck.
Your browser supports WebMCP — your agent can fill and send this form directly