In June 2026, Peter Steinberger reported that a production agent system spent $1,305,088.81 over 30 days while processing more than 603 billion tokens across roughly 100 Codex instances, operated by a team of three people. Oyedele Temitope argues that the central question was not only which model powered the workload, but what production harness kept hundreds of parallel agents coordinated after thousands of tool calls and decisions.

The article defines an AI agent harness as the runtime around a model that turns reasoning into execution. Its duties can include assembling instructions and context, managing tool access, maintaining persistent state, routing between models, verification, observability, guardrails and sandboxing (isolated execution environment). These responsibilities sit outside the model because they are systems engineering concerns rather than reasoning tasks.

Temitope says harness components usually appear after repeated production failures. Versioned system prompts reduce inconsistent behavior across runs; structured tool calls with retries prevent silent task failures; sandboxes limit risk when agent-generated code runs; durable storage preserves files and intermediate work across sessions; context management uses compaction, summarization and just-in-time context injection to avoid filling the context window; independent verification or tests catch errors before completion; human-in-the-loop approval handles irreversible actions; observability records traces, tool calls, cost and latency metrics.

The article cites Model Evaluation and Threat Research (METR) research on coding agents as evidence that complete agent systems matter. METR found Claude Code outperformed a simple Reason-Act (ReAct) loop 50.7 percent of the time, while Codex beat Triframe only 14.5 percent of the time. The article presents those results as a sign that a purpose-built harness can underperform a generic scaffold, and that the system wrapped around a model can reduce performance as well as improve it.

METR’s research also found that coding agents become less reliable as tasks grow longer and require sustained reasoning. The article describes this as a practical time horizon: reliability declines as task duration increases, with agents losing coherence, recovering less effectively from mistakes and struggling with extended execution.

The article contrasts METR’s findings with Anthropic’s engineering postmortem on Claude Code, where users reported lower coding quality even though the underlying model did not change. Anthropic traced the regressions to three harness-side changes: lower default reasoning effort to reduce latency, a context-management caching bug that repeatedly discarded prior reasoning after idle sessions and a system prompt change meant to reduce verbosity that unexpectedly reduced coding quality.

Temitope says the two sources measure different constraints. METR measures how reliably coding agents complete longer and more demanding tasks, while Anthropic measures how reasoning effort, context management and system instructions affect output quality while keeping the model constant. The article’s conclusion is that better models still matter, but production harnesses determine how reliably model capabilities become working outcomes.

In June 2026, Peter Steinberger reported that a production agent system spent $1,305,088.81 over 30 days while processing more than 603 billion tokens across roughly 100 Codex instances, operated by a team of three people. Oyedele Temitope argues that the central question was not only which model powered the workload, but what production harness kept hundreds of parallel agents coordinated after thousands of tool calls and decisions.

The article defines an AI agent harness as the runtime around a model that turns reasoning into execution. Its duties can include assembling instructions and context, managing tool access, maintaining persistent state, routing between models, verification, observability, guardrails and sandboxing (isolated execution environment). These responsibilities sit outside the model because they are systems engineering concerns rather than reasoning tasks.

Temitope says harness components usually appear after repeated production failures. Versioned system prompts reduce inconsistent behavior across runs; structured tool calls with retries prevent silent task failures; sandboxes limit risk when agent-generated code runs; durable storage preserves files and intermediate work across sessions; context management uses compaction, summarization and just-in-time context injection to avoid filling the context window; independent verification or tests catch errors before completion; human-in-the-loop approval handles irreversible actions; observability records traces, tool calls, cost and latency metrics.

The article cites Model Evaluation and Threat Research (METR) research on coding agents as evidence that complete agent systems matter. METR found Claude Code outperformed a simple Reason-Act (ReAct) loop 50.7 percent of the time, while Codex beat Triframe only 14.5 percent of the time. The article presents those results as a sign that a purpose-built harness can underperform a generic scaffold, and that the system wrapped around a model can reduce performance as well as improve it.

METR’s research also found that coding agents become less reliable as tasks grow longer and require sustained reasoning. The article describes this as a practical time horizon: reliability declines as task duration increases, with agents losing coherence, recovering less effectively from mistakes and struggling with extended execution.

The article contrasts METR’s findings with Anthropic’s engineering postmortem on Claude Code, where users reported lower coding quality even though the underlying model did not change. Anthropic traced the regressions to three harness-side changes: lower default reasoning effort to reduce latency, a context-management caching bug that repeatedly discarded prior reasoning after idle sessions and a system prompt change meant to reduce verbosity that unexpectedly reduced coding quality.

Temitope says the two sources measure different constraints. METR measures how reliably coding agents complete longer and more demanding tasks, while Anthropic measures how reasoning effort, context management and system instructions affect output quality while keeping the model constant. The article’s conclusion is that better models still matter, but production harnesses determine how reliably model capabilities become working outcomes.