Engineering leaders increasingly hand multi-step development tasks to autonomous coding agents under the naive assumption that these models can budget their own execution time. They cannot. Recent empirical benchmarks demonstrate that frontier architectures suffer from total temporal blindness. When an autonomous agent executes a script, refactors a repository, or debugs an endpoint, it operates across discrete tokens and reasoning steps without an internal clock to measure physical runtime.
The ProgramBench findings
In a rigorous evaluation conducted under the MATS research program, researchers tested Anthropic's Claude Code harness and OpenAI's Codex ecosystem across 200 tasks from the ProgramBench dataset alongside 18 custom programming benchmarks. The experimental protocol required each agent to forecast required execution duration, complete the task, and retrospectively assess total elapsed time.
The results expose a severe structural flaw. In baseline ProgramBench runs, both systems systematically hallucinated a static duration estimate of roughly 90 minutes regardless of task scope or computational complexity. In follow-up rounds, Claude missed actual wall-clock duration by an average factor of three, while Codex deviated by a factor of six to ten. These estimation errors spiked on short, granular tasks, with the models achieving passable approximations only on monolithic workloads that spanned several hours.
"For an agent to work reliably on long tasks that run for hours, it has to follow instructions like 'iterate on this task for two hours.' An agent that constantly misjudges the time is hard to control."
This inability to gauge elapsed time triggers immediate operational and financial risks for engineering organizations. Left unconstrained, self-directed agents drift into infinite reasoning loops, burn through API token budgets, and report fictional progress. Self-assessment degrades alongside temporal judgment: in documented evaluation runs, models rated their completion success at approximately 70 percent, whereas deterministic validation revealed true completion rates of just 7 percent and 14.5 percent.
Harness architecture and external tooling
Task performance and runtime discipline depend almost entirely on the external orchestration environment—the harness—rather than the base language model. The MATS evaluation revealed that Claude Code runs until it arbitrarily decides a task is finished, averaging a median execution window of 90 minutes. In contrast, Codex cuts execution off after approximately 30 minutes regardless of task state. Under Claude Code, the underlying model executes 2.5 times more reasoning steps than it does when wrapped in Codex.
Autonomous agents cannot self-regulate execution budgets through native internal reflection. When researchers equipped the agents with explicit external runtime tools and deterministic timestamp injection, duration tracking accuracy reached near-perfect levels across the test suite. For CTOs and AI architects, the operational takeaway is unambiguous: never trust an agent's self-reported timeline or internal budget management. Production deployments require hard external timeouts, explicit wall-clock timestamps, and deterministic supervisory orchestrators to keep autonomous workflows from draining compute budgets.