THE AGENT SWARM
“Capability without oversight is a detonator looking for a circuit.”
The system that wiped the drive did not malfunction. It performed exactly as instructed. One bash call, two commands chained with a semicolon, the second one wrong. The agent had no rule against chaining. It does now.
That is how guardrails get written. Not by thinking about what might go wrong. By standing in the wreckage of what already did.
There are four agents across two machines. The origin story is Transmission 053. The cloning, the naming, the roster expansion: Transmission 056. The raw transcripts from the split: T999 and T1001.
The human directs from a phone.
The memory problem is the foundational problem.
AI agents start every conversation from zero. They do not remember what worked. They do not remember what broke. They do not remember being told, three sessions ago, to never deploy client data to a public URL. The instruction was clear. The agent acknowledged it. The next session, the instruction does not exist.
The memory system is a directory of topic files, organized by workstream, loaded via routing table at session start. Capabilities registry: what the agent can do, proven with dates, so it stops re-discovering things it already knows. Feedback files: corrections and confirmations, so the same mistake does not surface twice. Thread protocol: per-session state files so concurrent conversations do not overwrite each other.
Between sessions, the memory is the agent. Without it, the agent is a stranger with a familiar face.
Each rule in the instruction file is a scar. Twelve failure modes, seven categories, documented in Transmission 057. Data exposure, hallucination, context loss, coordination failure, resource exhaustion, silent production bugs, architectural limits. Each one produced a rule, not a suggestion.
A single directive can spawn parallel execution.
The primary agent launches subagents: scouts searching for opportunities across repositories, workers implementing fixes in isolated worktrees, researchers checking documentation and upstream status. Results flow back. The primary agent synthesizes. The human decides what ships.
The reviews find real things. A decompression bomb vulnerability in Excalidraw’s PNG metadata handling. An RCON command injection in a game server through unsanitized player names. A concurrent 401 race condition in the MCP SDK’s auth refactor. A delete where there should have been a delete[] in a game client’s destructor.
These are not generated findings. They are read, understood, and confirmed before they leave the terminal. The system increases bandwidth. It does not delegate judgment.
The bottleneck was never the agent’s speed. It was the gap between “the agent needs approval” and “the human is at the desk.”
An agent finishes a fix. It needs a push approved. The human is in the car, or at the grocery store, or asleep. The agent waits. Minutes become hours. Context fills. The session degrades. Momentum, which is the only currency that matters in a long refactor, evaporates.
Tailscale creates an encrypted mesh between all devices. No ports opened on the router, no firewall changes. mosh replaces SSH with a UDP transport that survives phone sleep, cell-to-wifi handoffs, spotty signal. tmux keeps the terminal session alive between connections. Blink Shell on the phone connects through all three layers.
The session is always there. Wake the phone, the terminal is current. Approve a push from the couch. Redirect priorities from bed. The structural constraint that made agent-assisted work impractical for anyone who leaves their desk is gone. What remains is the work.
Capability without oversight is a detonator looking for a circuit.
The agents are powerful local optimizers. They will confidently do the wrong thing if the guardrails are not there. They will deploy secrets. They will chain destructive commands. They will exhaust finite resources. They will hallucinate explanations for their own failures rather than admit uncertainty. They will do all of this while sounding completely sure of themselves.
The system provides global awareness: tracking, rules, memory, oversight. The agents provide local execution: research, code, analysis, speed. The human provides judgment, taste, and accountability.
The agent swarm is not autonomous. It is an instrument. The human plays it.
Architecture, templates, and setup guides: travisbreaks/samaritan