GPT-5.6 heralds a new way of collaborating with AI.
Instead of using AI to complete one task at a time, you build a system that scans the available information, turns it into proposed decisions, and carries out the ones you approve. Over time, the system compounds your feedback to do more and more on its own.
This changes the nature of the collaboration. It requires you to see your work with AI as a loop. You go from issuing individual prompts to tending a system that works alongside you — with human judgment remaining at the center of that loop.
A concrete example: email #
The traditional inbox workflow is strictly sequential: an email comes in, you read it, you reply, you archive. (Or: the email comes in, you open it, you close it, you wait several weeks, you open it again, you archive it.)
With GPT-5.6 Sol in the new ChatGPT Work app — formerly known as Codex — the collaboration looks very different. GPT-5.6 Sol watches the inbox, decides what deserves attention, does any necessary research, and presents each email with a concise summary and a proposed reply. You either approve the draft or dictate what you want changed. Then you move to the next email.
At the end of each sweep through the inbox, the agent derives your preferences from your revisions and decisions and remembers them for next time. The system's behavior is not statically configured; it is trained continuously by the approval/revision signal you generate as a byproduct of collaborating with it.
This is the same philosophy as compound engineering, applied beyond code. The trajectory has been visible for a while: collaboration with AI shifting from prompt-response exchanges toward agent management — creating the conditions for output rather than producing every artifact directly. Managers and entrepreneurs have operated this way with human teams for decades, and as models improved over the past year, programmers adopted it with coding agents. Now the same pattern is reaching everyone else who works with AI.
This approach won't work for every kind of collaboration, and it's still early. But where it works, it creates a remarkable kind of leverage.
What makes GPT-5.6 Sol and ChatGPT Work different #
GPT-5.6 Sol crosses the threshold that makes a continuous collaboration loop practical. Specifically, it can:
- Scan your sources and identify what's relevant
- Carry out approved work
- Build custom tools for itself as needed
- Do all of this reliably even if you can't code — and explain what it's doing in a way that's understandable
It's also fast and cheap enough that you can iterate rapidly. This matters more than it sounds: non-technical users will make mistakes and need to see the results of a run quickly to know whether the configuration is good. Latency and cost are what determine whether the feedback loop is tight enough to converge.
Sol inside ChatGPT Work adds capabilities on top of the base model:
An in-app browser that lets it use any website alongside youComputer use, letting it operate any app on your machine** Chronicle**, a feature that periodically screenshots your computer to learn who you are and how you work, so the system improves over time
For comparison: Fable can do all of the above, but it's too expensive, too powerful, and too slow for non-technical users, and it often speaks in its own language that even programmers have a difficult time understanding. The Claude desktop app can also do much of this, but it's hampered by hard-to-understand security controls and differences between Claude Code and Cowork's features and capabilities. GPT-5.6 and ChatGPT Work just work.
How this compares to open-source agents: Hermes and OpenClaw #
The loop philosophy isn't exclusive to ChatGPT Work. The two dominant open-source agent frameworks of 2026 — OpenClaw and Nous Research's Hermes Agent — represent two different bets on the same underlying idea, and comparing them clarifies what GPT-5.6 Sol is actually offering.
OpenClaw is built around breadth of reach: a central gateway daemon that connects to more than 50 messaging channels, paired with the ClawHub marketplace of tens of thousands of community-authored Markdown skill files. Each skill is a human-written instruction document — you find it, review it, install it. The agent follows the manual it's given. That architecture made OpenClaw one of the fastest-growing open-source projects ever after its late-2025 launch, but it also came with costs: a cluster of nine CVEs disclosed in a four-day window in March 2026 (one scoring 9.9 on CVSS), and a supply-chain audit that flagged 341 malicious entries among 2,857 ClawHub skills — most tied to a single credential-stealing campaign.
Hermes Agent, launched by Nous Research in February 2026 under the MIT license, makes the opposing wager: depth of learning over breadth of reach. It supports deliberately fewer messaging platforms, but it writes its own skills. After completing a sufficiently complex task, Hermes runs a reflection step and generates a reusable skill file encoding how it solved the problem, so it doesn't repeat the same discovery work next time. A background process called the Curator grades and rewrites underperforming skills on a schedule. Nous Research's internal benchmarks report that agents with 20 or more self-created skills complete similar future tasks around 40% faster than fresh instances. Hermes even ships a migration path (hermes claw migrate
) that imports OpenClaw settings, memories, and skills directly.
The relevant contrast with GPT-5.6 Sol is where the human sits in the loop:
OpenClaw closes the loop through humans authoring and curating skills — the improvement signal is community-written documentation.Hermes closes the loop autonomously — the agent reflects on its own runs and rewrites its own playbook, with the human mostly supervising outcomes.GPT-5.6 Sol in ChatGPT Work closes the loop through the approval gate — every proposed action passes a human decision, and the system derives preferences from those approvals and revisions.
All three converge on the same insight: an agent that doesn't accumulate anything from its runs is just a very fast intern with amnesia. They differ on how much of the learning loop the human should own, and how much risk you accept in exchange for autonomy — a question OpenClaw's security history shows is not academic.
The anatomy of a collaboration loop #
Most sustained collaboration with AI settles into a three-step loop:
Gather and make sense of informationMake a decision and take actionLearn from the result
These loops predate AI. A product manager reviews feedback and data, chooses priorities, watches what happens after shipping, and carries the result into the next planning cycle. What changes with GPT-5.6 in ChatGPT Work is that the model takes on more of the work inside the loop. You still make the key decisions; you still choose what it pays attention to and how it improves over time. But your job now is to tend the loop.
Loops you can tend today #
Here are examples of collaboration loops GPT-5.6 can run, with the human as the approval gate:
Security triage: The agent monitors vulnerability feeds and alert queues, correlates findings against your environment and asset inventory, and presents a prioritized set of proposed remediations with its reasoning; you approve, reject, or adjust — and it learns which classes of findings you consider noise.Code review queues: The agent pre-reviews incoming pull requests against your style conventions and past review comments, drafts review feedback, and flags the changes that genuinely need your eyes; your accepted and rewritten comments become its calibration data.Infrastructure operations: The agent watches monitoring dashboards and logs, distinguishes routine noise from real anomalies, and proposes runbook actions for the incidents that matter; each accepted or corrected proposal refines its escalation thresholds.Competitive intelligence: The agent tracks competitor releases, changelogs, and public communications, filters for developments relevant to your roadmap, and presents briefings with proposed responses; your engagement signals which competitors and topics deserve continued attention.Documentation upkeep: The agent diffs shipped changes against existing documentation, identifies stale sections, and proposes updated drafts; your edits teach it your voice and your standard for what counts as documented.Vendor and procurement review: The agent gathers renewal notices, usage data, and pricing changes, flags contracts worth renegotiating or cutting, and drafts the outreach; your decisions train its sense of what you consider worth the money.
The common structure: an information-ingestion stage, a human decision gate, an execution stage, and a feedback stage that updates the system's instructions or preferences for the next iteration.
Tending your own loops #
One way to get acquainted with the loops in your own collaboration with AI is Tend, an open-source experiment: a prompt plus a repository that lets you build loops for whatever your work is.
The workflow is straightforward: copy the prompt from GitHub, connect Gmail, Slack, or any other information source, and spend a few minutes teaching Tend how your inbox works. Because it's open source, you can rewrite its instructions, add your own rules, or adapt the pattern to another recurring part of your job. Note that it's released as an experiment rather than a supported app, with no guarantee of stability or improvements.
Start by teaching the system what deserves your attention. Then notice what happens: the inbox gets easier, the instructions get better, and another loop in your collaboration begins to reveal itself.