Laude and MIT put persistent agency in 9,900 lines of Bash, and the bill arrives with it

Who decides how many model calls happen matters more than the line count. The main loop, called Thinker, repeatedly invokes a component named shellm, which asks a language model for reasoning, Bash commands or both, executes the commands, and continues until the model sets a FINAL environment variable ^[15]. There is no checklist unless the agent writes one, and each next thought is selected from earlier thoughts and new observations ^[13]. Iteration count is therefore a property of the agent, not of the request log ^[20], which is the mechanism under runtimewire's warning that the continuous loop can increase token spending ^[6]. Tiered compaction keeps recent events in full detail and summarizes older history ^[16], which shrinks the context per call and does nothing to the number of calls.

The shared stream has the same double character. Audel, the agent Laude has been running for several weeks, works as one agent over one stream reached through Slack, Telegram and a mobile app ^[8]. The behavior Laude puts forward as the point depends on that arrangement: following what different people are doing, returning to an old subject, contacting someone with no new prompt ^[9]. runtimewire flags the same stream as a cross-user privacy risk ^[7]. One mechanism, two names, and the team that installs it is choosing on behalf of everyone in the channel.

Shell access is not the delta. OpenAI's Codex CLI can read, modify and run code locally ^[17], and GitHub Copilot CLI can execute shell commands subject to permissions or approval ^[18]. An approval prompt assumes a human is present at the moment the command runs. Laude's clearest example, dated August 5th, happened while nobody was interacting with the agent: it checked a background recall process it had built, found the code reading an environment variable that was never set, searched the codebase to confirm, rewrote the process to read from a pipe, verified the repair, and the fix was merged, with the log putting the sequence at 48 minutes ^[12]. Attendance is what changed, not privilege.

What stands in for evaluation is that log. More than 50 commits from Audel's fork were pulled into the main repository ^[11]. On its first day the agent audited eight stale Git branches and came back 10 minutes later to correct its own count ^[10]. There are no independent performance results, and Laude says it judges changes to Audel mainly by qualitative observation ^[14], so every quantified figure on offer is first-party ^[19]. Laude's grantmaking is anchored by Konwinski's $100 million pledge ^[5], which is part of why the artifact turned up as Apache-2.0 code on GitHub instead of a product with a score attached ^[3].

The compensating handle is instrumentation. Thoughts and actions are stored in a directed graph of JSONL files that supports forks and merges ^[15], and the core is 9,900 lines of Bash in bin/ and thinkers/ ^[4]. Absent an eval, that graph is the only way to reconstruct what an unattended agent spent and what it touched, after the fact.