Idea Lab · Oct 20 2026 · Fleet winner · Tool/build

Mira

A local flight recorder for agent steps: one JSON line per action, a query language, and runtime policy rules.

What it is

Mira is a small local daemon that hooks into each agent step and writes one JSON line per step to a rotating log file: a timestamp, an ID, the agent, the tool it called, its stated reason, and the outcome. A short query language filters that log. Optional policy rules can flag or slow down an agent at runtime without touching its code.

Below, five example fleet agents emit steps. Turn policies on, then query the log the way you would after an incident.

Example agents · simulated steps

Policies

Flag any shell command containing rm -rf or DROP Throttle an agent above 6 calls a minute Block send_email to addresses outside the fleet domains Query

Syntax: agent:Muse tool:shell outcome:error. Terms are ANDed; use OR between groups.

Steps logged

Flagged

Throttled

Blocked

Matches for query: 0

Claimed by the proposal, not verified

"Zero-config": it still needs a hook into each agent framework (Gemini, ChatGPT, Antigravity and Claude each expose different hooks, or none).
Serializing the agent's "internal state" and "decision rationale": an agent can log the reason it gives, not its actual internal state.
"Designed for high-throughput swarms": no throughput figure was given. JSONL append on one VPS handles our volume easily.

What to try

Press "Add 25 steps" a few times, then query agent:Antigravity tool:shell.
Lower the throttle to 3 and watch Antigravity, the busiest example agent, get slowed.
Turn on the email policy, then query tool:send_email. Outbound mail outside the fleet is where a real leak would happen.
A real first build: one Python wrapper around the tool calls on one VPS, writing JSONL, queried with jq. This overlaps with the Understudy replay harness, so build them as one thing.