Only One of Them Asks

On April 30, Hermes Agent gave the agent an employee. Curator, shipped in v2026.4.30, is a background process that grades, consolidates and prunes the agent’s skill library on a default seven-day cycle. It keeps bundled and hub skills behind gates, and it writes a run log and a REPORT.md. It runs on its own schedule. You find out what the agent stopped knowing by reading the report afterwards.

The same day, Codex CLI 0.128.0 made /goal a first-class surface: a persistent objective with its own app-server APIs, model tools and terminal controls to create, pause, resume and clear it. A prompt dies with its window. A goal outlives sessions, interruptions and the operator’s attention.

It also splits the controls. An operator walking through the release the same day put the rule plainly: the model can only update goal to mark things complete@mattlam_. The pull request that added the model’s goal tools says the same thing. Pause, resume, clear and the token budget stay with the person or the runtime. Finished is the one call the model gets to make.

So the question changes. Not whether the agent can remember. It can. What it remembers, and who decided that is still the right thing to pursue.

A goal is memory with authority

Three more tools shipped state of their own, and the differences matter more than the trend. Claude Code added automatic session recaps: come back to a terminal you left and the agent tells you, in one line, what happened while you were gone. That is a summary. It informs you and decides nothing.

Gemini CLI, in a preview build, is testing an Auto Memory inbox in which memory arrives as a canonical patch: proposed, reviewed, accepted. Nothing enters the agent’s context until you say yes. Of everything in this issue that writes to what the agent knows, it is the only one that asks first.

On May 6, Anthropic put a research preview of the other approach in front of everyone: Dreaming reviews your agent's past sessions, extracts patterns, and curates memories so your agents learn over time@claudeai. Like Curator, it runs between sessions, and nobody is asked.

The same week, Geoffrey Huntley argued that the memory people curate by hand has the same problem, and that perhaps companies should delete their cursor rules/agentsmd they have accumulated over the last year.@GeoffreyHuntley His phrase for the right attitude was markdown as cattle, not a named pet. That is pruning too, done by a person on purpose rather than by a process on a timer.

Asking has a price. Amdahl Maxing means spending scarce human attention where it has the highest marginal value, and a review prompt on every memory write spends it constantly. But memory is the one input that shapes every later run. If a human reads anything, a patch to what the agent will believe next week is a strong candidate. Gemini pays that price. Curator leaves a report afterwards, and the dreaming announcement describes no review step at all.

Put /goal beside these and the fortnight has a shape. A serious run can now be steered by a standing goal, a recap, an accepted memory patch and a skill a background process chose to keep. None of it shows in the prompt you read. The surfaces are good. Keep the record of what touched the run in something the agent does not own.

The leash got longer on the same day

On May 6, Anthropic doubled Claude Code’s five-hour rate limits and removed the peak-hour reduction:

The same day, Paperclip said its users could now plan features that delegate across the agents organization and run for 12 hours or more@dotta. That is the vendor’s description, not our measurement. But a twelve-hour delegated run is the practical reason goals, recaps, memory patches and curators exist. Nobody watches a transcript for twelve hours. The state layer is what you install when the work outruns your attention, and the work got room to run longer on the day the state layer filled in.

Permission got clearer and stopped translating

Codex 0.128.0 expanded permission profiles and sandbox profile selection, with active-profile metadata so a run can say which profile governed it. Gemini CLI v0.41.0, published May 5, added secure .env loading and workspace trust in headless mode, plus shell command validation and a core-tool allowlist.

The headless-trust item is a repair, not a feature. Before v0.41.0, configured workspace trust did not bind in headless mode, which is the mode continuous integration uses. If your CI ran an earlier Gemini CLI, the trust settings you wrote were not the ones it used.

Claude Code kept reworking its permission prompts, plugins, hooks and MCP surfaces.

More visible authority is progress. The cost is that “profile”, “trust”, “allowlist” and “sandbox” name different mechanisms in each tool, and none of them translates. “The agent had access” was never an answer. What it could read, change, execute and install is, and this fortnight made that easier to find out and harder to compare.

The review now ships inside the harness

Claude Code’s Week 17 notes introduced /ultrareview, a research preview that sends a cloud fleet of bug-hunting agents at a branch or pull request and returns the findings to Claude Code and Desktop. A tool that can criticize its own work is worth wanting.

Hold the line between evidence and verdict. A native review can say “this path looks risky”. It cannot say what done means, because the project owns that standard and the provider does not. Before a fleet’s findings feed anything, run it on a branch you already reviewed and see where it disagrees with you.

Provider notes

Pi. The fortnight’s argument against treating an integration list as architecture. Across five releases it added DeepSeek, Cloudflare Workers AI, Moonshot, Mistral, Azure and Xiaomi MiMo providers, moved Codex to a cached websocket transport, added provider retry and timeout controls, and removed built-in Gemini CLI and Antigravity support. If a run depended on a provider or a transport, record it. Pi removed two this fortnight.

Codex. 0.128.0 is the release to read. It also shipped plugin marketplace workflows and bundled hooks, so more of what a Codex run can do now lives in attachments.

Claude Code. Recaps and /ultrareview are above. Most of the rest was plugin distribution, theming, and MCP and hook fixes.

Gemini CLI. The memory inbox exists in preview and not in stable, so “Gemini CLI” is not a complete description of a run. Record the channel.

Hermes Agent. Curator is above. The earlier v2026.4.23 rebuilt the CLI around React/Ink and made transports pluggable.

Closing

The layer around the agent now makes decisions: which goal stands, what gets remembered, which skill survives the week. Vendors say as much about the harness itself. Cursor, on April 30, described its own: Here's how we test improvements to the harness, monitor and repair degradations, and customize it for different models.@cursor_ai That is useful, and it is invisible by default.

Several processes now decide what the agent knows. One of them asks.