Your AI knows the code. Who knows the plan?

I A brilliant colleague with a patchy memory

Picture a coding agent asked to fix login. Twenty minutes later it has redesigned authentication, hired three copies of itself, and rediscovered yesterday's failed fix. Everybody is busy. Login is still broken.

The missing skill is judgment about the work itself: which evidence matters, what has already failed, whether to continue, and when another agent would help. Functional self-awareness becomes useful at exactly those decisions.

The Situational Awareness Dataset gives this idea measurable parts: predicting one's behavior, identifying one's circumstances, and following instructions that depend on self-knowledge. Across its 16 tested models, results exceeded chance but still left substantial gaps. General-knowledge scores explained only part of the variation.^[1]

A more revealing test puts that knowledge to work. Christopher Ackerman's metacognition study asks models to use confidence and predictions about their own answers strategically. It finds increasing abilities alongside coarse resolution and strong context dependence—even between similarly capable models.^[2] Knowing an answer and knowing how to manage the next step are separable skills.

II The context window has two different failure modes

First, information can remain in the input yet fail to influence the answer. Lost in the Middle found that moving relevant material into the middle of a long context degraded performance on its question-answering and retrieval tasks, including for long-context models.^[3]

Second, information can disappear from the working history. Truncation removes old material; compaction replaces it with a shorter account. If that account preserves “tried a cache fix” but loses why it failed, the next attempt starts with an incomplete map.

OpenAI demonstrated the cost with GPT-5.6 Sol: preserving reasoning and enabling compaction changed its ARC-AGI-3 public-set score from 13.3% to 38.3%, while using roughly six times fewer output tokens. The original harness discarded reasoning after each action and eventually truncated earlier moves.^[4] Same model; a very different ability to build on its work.

The score is Relative Human Action Efficiency on that public task set. The engineering lesson travels further: memory handling belongs in the performance investigation.

III Give the team different jobs

Parallel agents buy separate working contexts. That helps when one worker traces a dependency, another inspects a browser failure, and a third reviews the proposed patch. Copying the same incomplete brief to all three spreads the same blind spot.

Anthropic's research system found its strongest fit in independent, breadth-first investigation. Its early agents also produced duplicate searches, endless exploration and excessive spawning. The production write-up reports about 15 times the token use of chats for multi-agent systems in its data.^[5] Delegation needs a reason worth paying for.

For the login example, keep one writer. Give the reviewer the exact diff and failing test, and let a separate investigator inspect the original error. Each returns evidence the writer can use. This is a suggested division of work; closely coupled edits still meet at an integration check.

IV Keeping useful context through a handoff

A harness runs the model's working loop: tools, conversation state, context handling and execution. OpenAI's ARC experiment shows why those choices matter. With multiple harnesses, preserve their useful native behavior and share explicit work products between them: source revisions, diffs, evidence and decisions.

For our login repair, keep the task, working branch and review together. Let the selected harness execute with its configured permissions. Switching a provider then becomes a controlled handoff: the new worker gets the failed test and current patch instead of an invitation to start the project again.

The resulting architecture supports judgment while making its consequences inspectable. Astra pushes the model side of that arrangement forward.