The request log: one task board for all our agents
In short: the tower runs three artificial intelligence agents: Claude Code, opencode and Antigravity (Google's agent). Each had its own way of tracking — or forgetting — what it was asked. Now they all write in the same log, with no way to opt out, and the human sets the priorities.
1. The problem
An agent gets a request, then another while it works, then a third. It answers the last one. The others vanish. Multiply by three agents, and nobody knows who is doing what, or what was forgotten. Writing “forget nothing” in the instructions is not enough: the agent reads it, then forgets that too.
2. The principle: a gate, not an instruction
The tower applies the same idea everywhere: a rule that matters must be held by code, not by the agent's good will. Each agent can run small programs at precise moments: when a message arrives, before it uses a tool, when it wants to stop. These are called hooks. The log hooks into those three moments.
3. What the log does, in four moves
Record. Every message from the human becomes a task, automatically, before the agent even reads it. A priority is guessed: “at the end” gives 5, “urgent” gives 1, otherwise 3.
Remind. On every turn, the list of open tasks is put back in the agent's head, most urgent first.
Keep the order. The log refuses to let an agent mark a less urgent task “done” while a more urgent one is waiting. It must finish it, or say why it is blocked.
Check before stopping. The agent cannot hand back control if a request was never filed, and its answer must say how many tasks remain open.
4. Priority belongs to the human
The log shows up in the cockpit, on a column board: to do, in progress, blocked, later, done. Each card has a priority menu, from 1 (urgent) to 5 (at the end). When Patrick changes a priority, it is marked as his, and no agent can change it again: the log refuses. An agent trying to pass itself off as him is stopped before it even runs its command.
5. One engine, three plugs
The core is a single small program, shared by all. Around it, three adapters, because each agent speaks a different language: Claude Code has its hooks, opencode its plugins, Antigravity its own hooks file. No new server was added: the board is served by the cockpit that was already running.
6. What the real trials taught us
The sandbox tests were all green. The real trials found three failures they could not see: opencode refused to load the plugin because it exported something other than a plugin; pasted messages started with a character the log mistook for a technical line; and Antigravity had been blocking all its tools since the morning, because of two old guardrails — one damaged mid-write, the other answering “nothing” where the agent expects “allowed”. Each failure now has its own test.
In one sentence: an agent no longer has to remember what it was asked. The log remembers for it, makes it follow the order, and the human decides the order.