Daily Digest — 2026-07-15

Addy Osmani's keynote at the AI Engineer World's Fair highlights a fundamental shift in the role of software engineers as AI agents become more prevalent. The central theme is that the engineer of the future will be defined not by their ability to execute tasks, but by their capacity to choose what is worth doing, and then own the evidence, understanding, and verdict for work increasingly automated by agents.

Osmani clarifies three crucial terms: Quality, which is the system of checks that produces evidence; Verdict, the human/accountable decision made from that evidence; and Answerability, the ability to explain and stand behind the verdict later. This framework underscores that while AI improves efficiency, human oversight and accountability become more critical. The speaker cites Boris Cherny's five archetypes for future roles: Prototyper, Builder, Sweeper, Grower, and Maintainer. He emphasizes that roles are unbundling from traditional craft boundaries and rebundling around ownership. The title matters less than the part of the system an engineer can genuinely own.

The talk introduces Harness Engineering and Loop Engineering as key architectural concepts. Harness engineering involves building the scaffolding (prompts, tools, state, constraints, feedback loops) around an AI model to turn it into an agent that can reason and decide. Loop engineering extends this by designing systems where agents continuously prompt, check, remember, and decide what happens next, moving beyond one-shot prompting. These loops enable agents to act like infrastructure, handling tasks recursively until done, thereby changing the work but not eliminating the engineer.

Osmani presents an "Agentic Software Factory" model where the agent inner loop (guide/context, generate, verify/solve) processes product intent, incidents, and user feedback, producing evidence (tests, diff summaries, risk notes). The engineer outer loop then provides the "human verdict" (ship, block, or redirect). The speaker asserts that the win is not removing people from the loop but moving human judgment to the highest leverage checkpoint. He backs this with data from Sonar's 2026 survey, showing that 42% of committed code is already AI-generated or assisted, with projections for 65% by 2027. This indicates that AI code share is no longer marginal; it's entering the commit history, making answerability an engineering requirement rather than a philosophical concept.

Clean code is also highlighted as beneficial for agents, leading to 7-8% fewer tokens and 34% fewer file revisits with the same pass rate. However, the talk also addresses the "Trust without Capacity" problem: 96% of engineers don't fully trust AI-generated code, and 38% say reviewing AI code takes longer than human code. This reveals a trust gap where skepticism is high, but verification is not keeping up. The "Governance Gap" data shows that 92% of organizations report some governance challenge with AI-generated code, with review/validation being a bottleneck (85%). Adoption of AI is moving faster than policy (80%), and many cannot reliably distinguish AI vs. human code (43%). This argues for a "Human Verdict" interface with first-class provenance, intent, and ownership.

Osmani introduces two terms for career development in the AI era: Alpha and Decay. Alpha is the gap between what you can do meaningfully better than current models, and Decay is how fast models catch up to that alpha. He posits that while capabilities like speed, recall, and verification will decay rapidly (with a half-life of one release), taste (judgment before objective metrics exist) and judgment itself will decay much slower. Taste is hard to create but easy to copy, making it valuable as it dictates what others copy next. The future value for engineers lies in discerning which options deserve to exist, moving beyond simply executing tasks.

Engineers must actively avoid three pitfalls to remain effective and accountable: 1. Cognitive debt: The erosion of understanding and memory around problem-solving due to deferring more to AI. Learning through AI-generated code leads to lower comprehension (17% lower, per Anthropic's study). When tasks become long and parallel, the scarce resource is no longer generation but answerable delegation. 2. Cognitive surrender: Blindly accepting what AI gives and stopping critical thinking. A Wharton study showed that 73% accepted the wrong AI answer and felt more sure about it, highlighting that the failure mode isn't using AI, but borrowed confidence. 3. Orchestration tax: Diminishing returns and cognitive drain from managing parallel AI agents. More agents don't increase human cognitive bandwidth; instead, intentional attention allocation becomes crucial for managing loops.

Accountability is not what remains after agents get good; it's what enables the rest of the system to scale. Engineers must own the outer loop, making decisions, verifying evidence, approving, and carrying the consequences. The operational rule is "Explain it or don't ship it," because you cannot answer for what you cannot understand. Automation moves the floor, shifting work from automated task layers (typing, boilerplate, routine fixes) to higher-level engineering: Loop Design, Evidence Design, and Brownfield Care. This means not less engineering, but more Surface Area requiring taste, verification, ownership, and care. The enduring takeaway is that every time software development became easier, we ended up writing exponentially more of it. The future belongs to engineers who make agent work legible, verifiable, and worth shipping.

The video delves into OpenAI's GPT-5.6 Ultra reasoning level within Codex, highlighting that it is fundamentally misunderstood and mislabeled. The author argues that Ultra isn't a new reasoning level but rather a 'skill' or toggle that forces the underlying agent and its subagents to operate at Max reasoning, leading to significantly increased token consumption and costs without necessarily providing a proportionate benefit in output quality.

Initially, the video shows how OpenAI's documentation and the Codex desktop application UI present Ultra as a distinct, higher reasoning level. However, the author demonstrates that in the Claude Code environment (Anthropic's tool, which OpenAI appears to have 'borrowed' the concept from), Ultra is clearly labeled as 'ultracode' alongside 'xhigh + workflows,' indicating it's a workflow-triggering mechanism rather than a reasoning level. When Ultra is selected in Codex, the base reasoning level defaults to 'Max' rather than being an independent setting, which the author states burns tokens at an excessive rate for only a minor performance improvement (4-10% better at 2x the cost).

The core problem identified is that when Ultra is enabled, both the parent agent and all subsequent subagents operate at Max reasoning. This leads to a massive token burn because the subagents are also running at the highest, most expensive reasoning level. Furthermore, the current version of subagents used by Sol (OpenAI's model) does not allow for granular selection of reasoning levels for individual subagents. This means that if the parent agent is set to Ultra, all its spawned subagents will also run at Max, making cost optimization impossible.

The video then contrasts V1 and V2 subagent implementations. V1 (Worker Pool) uses a dispatcher to hire temporary helper threads by ticket number, where the parent spawns helpers, receives opaque thread IDs, waits on chosen IDs, and manually closes helpers when their slots are no longer needed. Each helper operates with its own independent context, reducing pollution. V2 (Task Tree), which is under development and is the default for new Sol models, is presented as a named project team with an org chart and mailboxes. In V2, children agents can create their own children, and teammates can message one another, with finished work flowing back to the parent's mailbox. However, this implementation has significant flaws. By default, V2 shares the full eligible history of the main thread with all subagents, leading to context pollution and unnecessary token usage. Additionally, it filters out tool calls, which further breaks caching mechanisms and increases costs. The nesting capabilities in V2, combined with default full context sharing, create a recursive loop where costs can skyrocket. The author describes a personal experience of hitting a 5-hour usage limit twice in under an hour due to this behavior.

The author attributes this problematic design to OpenAI attempting to copy Anthropic's 'ultracode' concept but implementing it poorly. He stresses that workflows in Claude Code are a more robust solution for programmatic control over subagents. In Claude Code, workflows allow for defining specific phases with typed outputs and dynamically creating subagents with tailored instructions and controlled context sharing. This programmatic approach gives engineers precise control over resource usage and prevents the unchecked token burn seen in Codex's Ultra mode.

Ultimately, the video advises users to avoid using Ultra in Codex for now. It encourages users to 'just wait' for OpenAI to refine its implementation, as the team has shown responsiveness to feedback. The author believes that OpenAI needs to rethink their subagent strategy and provide more granular control, similar to Claude Code's programmatic workflows, to enable efficient and cost-effective use of powerful AI models.