Every rig, instantly filterable.
Claude manages, Codex grinds
steal → Use one model to plan architecture and another to churn through build steps.
The agent never sees the API key
steal → Inject credentials at an egress proxy so a compromised agent can't leak them.
Private data stays in Hermes; cloud chores go to Grok Bot
steal → Draw a hard line between your local agent and your cloud agent, and you approve what crosses.
One task, one thread, so bad corrections die with it
steal → Start a fresh thread per task; carry over only the memory you mean to.
Five harnesses, three machines, one orchestrator
steal → Install several harnesses on worker machines and drive them from one place.
Route by task type and make each agent review the other
steal → Have the agent that didn't write the change review it.
Three desktop apps out, one terminal multiplexer in
steal → Use one session view for all CLIs instead of one app per harness.
Every model gets a written job: driver, workhorse, gatekeeper
steal → Write each model's role into AGENTS.md so routing isn't in your head.
Broke up with Claude Code: open-weight models at about two cents a task
steal → Put every model call behind one gateway with per-app budgets, and force a test run before 'done'.
Ticket in, PR out: four skills and a reviewer in CI
steal → Split the pipeline into ticket, plan, build and address-review skills.
22 MCP servers, 300+ tools, two used daily
steal → List your MCPs next to what you used this week. Disconnect the idle ones.
A PM, a developer and a designer bot for each area of life
steal → Give bots a shared task board (Linear/Notion) so they work from the same queue.
OpenCode on DeepSeek V4.1 Flash or GLM-5.3 instead of a native seat
steal → Point your harness at a cheaper model endpoint and test it on your own tasks.
Laptop as control room, home PC as the engine
steal → Run agents on an always-on box behind one model gateway; reach it with zero open ports.
Your skills folder is a graveyard. Measure it.
steal → Track which skills actually fire, then delete the rest.
Anthropic cut over 80% of a harness system prompt for newer models
steal → Re-read your system prompt after each model upgrade; newer models need less.
Full YOLO, but on a disposable VM
steal → Run with approvals off only on a machine you can wipe.
103 capabilities installed, 90 never fired in 30 days
steal → Mine your session logs for what actually ran, then cut the rest.
A slimmer .cursor: thin always-on rules, glob-scoped detail
steal → Move catalogs out of always-on files into a reference folder.
Hermes coordinates, Cursor builds, local models for sensitive work
steal → Use a persistent agent as the coordinator and delegate code to your IDE agent.
One config folder, symlinked into three agents
steal → Keep one AGENTS.md and skills folder; symlink it into every harness.
OpenClaw and Hermes in separate VMs, so neither is a single point of failure
steal → Put each agent in its own snapshot-able VM; fix credentials next.
OpenCode + OMO, 100% local
steal → Pair a local agent with a self-hosted forge like Gitea and nothing leaves the box.
Cheap heartbeat model, API fallbacks, aliases for /model
steal → Run cron/heartbeat on a cheap model and keep the strong one for real work.
Every related repo in one Cursor workspace
steal → Open sibling repos as one multi-root workspace and delete ceremony ruthlessly.
Cowork with a memory folder and read-only connectors
steal → Start every connector read-only; turn on write access one at a time.
Every MCP ships disabled; enable one or two per session
steal → Set MCP servers to off by default and switch them on only for the task.
Change one edit tool, watch a weak model jump from 6.7% to 68.3%
steal → If a cheap model fails at edits, fix the edit format before you upgrade the model.
Each OMO agent gets its own model
steal → Route by role: put the expensive model only where reasoning matters.
Let the agent write its own skills instead of installing MCPs
steal → Ask your agent to write a small skill for the job before reaching for an MCP.
Left Claude Code for Pi: smaller prompt, any model
steal → If you switch models often, a model-agnostic minimal harness pays off.
Claude Code's creator runs it almost stock
steal → Check one shared CLAUDE.md into git and start in Plan mode before you add anything else.
PreToolUse hooks as a seatbelt: no rm -rf, no push to main
steal → Write one PreToolUse hook that blocks rm -rf and direct pushes to main.
Four tools. No MCP. No sub-agents. It ships.
steal → Before adding a tool, ask if read, write, edit and bash already cover it.
A dedicated repo where research agents get the open internet
steal → Give risky, network-hungry research its own repo instead of loosening your main one.
Parallel agents from throwaway /tmp checkouts, no worktrees
steal → For quick parallel tasks, a fresh clone in /tmp beats worktree bookkeeping.
The maximalist: 18 agents, 50+ skills, and the advice to start with three
steal → Start with CLAUDE.md plus two or three skills; grow only from real friction.
CLAUDE.md under 30 lines, detail pushed to path-scoped rules
steal → Keep the always-on file tiny and scope everything else by path.
One of each, then stop
steal → Ship with one CLAUDE.md, one MCP, one hook, one skill. Add the second only when the first hurts.
MCPs scoped per project, not per machine
steal → Install MCPs at project scope unless you truly need them everywhere.
3–8 Codex agents, one folder, one 800-line AGENTS.md
steal → Have parallel agents make small atomic commits in one folder; swap MCPs for CLIs.
A full OpenCode agent team on free models
steal → Give every agent a fallback array so a rate limit never stops the run.
DeepSeek's own harness has a bash + editor mode
steal → Benchmark your model with the smallest toolset first.
One OpenClaw agent 24/7 plus one research subagent
steal → Start with one main agent and one helper, not a swarm.
ClawHub off, third-party skills rewritten locally
steal → Read and rewrite any third-party skill before it runs on your machine.
Three Hermes bots in Docker, one Telegram topic each
steal → Give each bot one domain and mount only the folders it needs.
Short GEMINI.md, workflows as skills
steal → Move repeatable workflows out of the main instructions file into skills.
Do it once, save it as a skill, then schedule it
steal → Only automate a task after it has worked by hand.
One ~/.ai folder feeds every coding agent
steal → Keep shared rules, security policy and MCP list in one home folder all agents read.
Pi cold start: lazy-load the heaviest tools
steal → Make your heaviest tools opt-in so their schema is paid for only when used.