Skip to content

Knowledge base

Human accountability

What stays with the engineer when agents write the code: delegate, review and own, the new middle loop, and three manifestos that put it in writing.

Written by

Updated:

Agentic Engineering 5

Our position. We delegate, review and own (opens in a new tab): the agent does the mechanical work, and the engineer reviews it and keeps ownership of the decision. Delegating to an agent is closer to delegating to a junior engineer than to handing off to a contractor. You set the context, you read the work, and you stay accountable for what shipped.

The boundary reshapes the engineer’s day. Review becomes the main task, a new middle loop checks each agent proposal against the right problem, and engineers keep enough hands-on coding to stay able to supervise.

What stays with the human when agents write the code?

None of this removes the human from the hook. Three manifestos make the accountability explicit: the AI Manifesto’s five principles say never let a model speak, think, or own your work, replace your curiosity, or discourage someone else (The AI Manifesto (opens in a new tab)); the AI-Augmented Software Development Manifesto reworks the Agile Manifesto for agents, valuing human taste over autopilot, tight feedback loops over long unattended runs, tested software over abundant output, and clear adaptable intent over both rigid specs and vibe coding, on the principle that code is cheap and understanding is expensive (AI-Augmented Software Development Manifesto (opens in a new tab)); and the Outcome Engineering (o16g) manifesto pushes the frame further, that it was never about the code, verified reality is the only truth (code is a vanity metric, vibes are not tests), the backlog is dead so you manage to cost not capacity, and because models drift you continuously audit the agent against the domain and make risk a blocking gate that stops the line (The o16g Manifesto (opens in a new tab)).

The skills that hold their value shift accordingly: version control as the coordination backbone between humans and agents, requirements decomposition, testing, system design, operations, and code review, which becomes the most time-consuming task of the day, while syntax recall and typing speed lose value (DoltHub – The Agentic Software Engineer (opens in a new tab), 2 July 2025). Code review’s deeper function survives the shift: it is mental alignment, keeping the team’s model of the system in sync, which is why reviewing the plan matters as much as reviewing the diff (Advanced Context Engineering for Coding Agents (opens in a new tab), 2 December 2025).

Thoughtworks gives the same shift a structural name. The old inner loop (typing, compiling, and fixing syntax in the IDE) and outer loop (everything after git push: review, CI/CD, deploy) have been joined by a new middle loop that triggers the moment an agent proposes a solution, where the human checks whether it solved the right problem, guards against architectural drift, and synthesizes the parallel output of several agents into one coherent system. The discipline it names, supervisory engineering, has three pillars: directing (breaking a system into agent-sized chunks and codifying standards so the agent does not hallucinate its own patterns), evaluating (reading plausible, beautifully indented code and asking whether it catches the edge cases, uses a deprecated API, or is confident nonsense), and correcting (stitching several agents’ codebases into architectural coherence).

The habit change it points at is that you used to slow down while writing code, and now you have to force yourself to slow down while reading, questioning, and auditing it (Thoughtworks – Supervisory Engineering: Orchestrating Software’s Middle Loop (opens in a new tab), 3 June 2026).

The higher-leverage move is to stop reviewing every diff by hand and start training the system, extracting each lesson into a durable rule or skill file so the next run is better rather than fixing the same class of error in chat (How I Use AI to Code (opens in a new tab), updated 11 September 2026). The counterweight is the supervision paradox: supervising an agent well requires the coding skills that atrophy when you delegate to it, and skill drop-off is already measurable, so a defensible discipline is to use AI for specs, plans, and research while keeping hands on implementation and never generating more than you can review in a sitting (Agentic Coding is a Trap (opens in a new tab)). Even inside a spec-driven flow on a genuinely hard problem, the reliable ending is sometimes to throw the agent’s plausible-but-wrong version away and rewrite the core by hand once you understand it well enough (Partner with the AI, throw away the code (opens in a new tab), 31 July 2025).

The same lesson shows up as a precondition, not only an ending. A Netflix account of a stalled authorization refactor, where permission checks were woven through business logic and the agent spiraled every time it tried to untangle them, resolved only after the team did one migration entirely by hand, no AI, reading the code until they knew which invariants had to hold, then fed that pull request back as the seed for the agent’s research phase; the research-plan-implement loop worked because they earned the understanding first, and no better prompt or model substitutes for it (The Infinite Software Crisis (opens in a new tab), 20 December 2025).

The deeper diagnosis is that AI is the ultimate “easy” button and easy is not simple (the distinction: simple is unentangled structure, easy is proximity), so agents preserve accidental complexity as readily as essential, treating five-year-old technical debt and load-bearing logic as equal patterns to keep; skipping the thinking to match generation speed also atrophies the engineer’s own instinct for when a system is turning dangerously complex (The Infinite Software Crisis (opens in a new tab), 20 December 2025).

One practitioner talk sharpens answerability into an evidence package and an attributable production verdict. The agent returns task-appropriate diffs, tests, logs, rationale, traces, or screenshots. A person judges whether that evidence is sufficient, then decides whether to ship, block, redirect, or accept the risk (The Engineer of the Future Chooses What Is Worth Doing (opens in a new tab), 14 July 2026, practitioner synthesis).

It separates capability from agency. Agents can run the inner execution loop, including investigation, implementation, tests, and reporting. Humans own the outer loop of deciding, verifying, approving, and carrying the consequences. More parallel agents raise the orchestration tax because human attention does not parallelize with them (The Engineer of the Future Chooses What Is Worth Doing (opens in a new tab), 14 July 2026, practitioner synthesis).

Its career frame calls today’s advantage over a model “alpha” and the model’s progress “decay.” The terms are practitioner constructs, but the action is concrete: durable judgment leaves examples and reasons a team can reuse, while cognitive surrender begins when the agent’s answer becomes the person’s answer before they form their own (The Engineer of the Future Chooses What Is Worth Doing (opens in a new tab), 14 July 2026, practitioner synthesis).

An economics paper summarized in the 2025 Work Trend Index gives three reasons human work persists as AI capability grows: a human and an AI working in tandem can beat either alone, people prefer a human for some things, and some decisions carry moral weight that society expects a person to answer for (“What Will Remain for People to Do?”, Knight First Amendment Institute, April 2025, as reported in Microsoft - 2025: The year the Frontier Firm is born (opens in a new tab), April 2025, vendor). The report works the three through examples. AI optimizes delivery routes while human dispatchers handle real disruptions, AI models risk while clients still want a trusted advisor for a large decision, and AI summarizes case law while a high-stakes negotiation stays with a lawyer. Only the first limit erodes as models improve. The third is why an accountability boundary is a standing design decision rather than a workaround for weak models.

Answerability needs a record of which lines an agent wrote, and a vendor-neutral format for that record appeared in January 2026. Agent Trace, a CC BY 4.0 specification at version 0.1.0 and RFC status, is backed by Cognition, Cursor, Cloudflare, Vercel, Google Jules, Amp, OpenCode and git-ai. It attributes each file and line range to a contributor typed as human, ai, mixed or unknown, names the model, and links the conversation that produced the change by URL instead of copying it into the commit, with optional content hashes (Agent Trace (opens in a new tab)). It leaves storage open, whether local files, git notes or a database, and states plainly that it does not track legal ownership, training-data provenance or code quality. Cognition’s announcement adds the motive: an agent that can retrieve the reasoning behind existing lines should stop re-deriving and reinventing it (Cognition - Agent Trace (opens in a new tab), vendor).

Resources

Human accountability