On August 16, the long agent trajectory became the main engineering object. Prime Intellect gave agents up to eight days of scientific search. HyMem split the plan from detailed execution traces. Bounded Agents moved delegated authority through a separate chain, while code review forced reviewer and critic to disagree.
Even economics now changes within a day: DeepSeek's new rates took effect at exactly 16:00 UTC. For a long-running agent, model version, reasoning, price, and authority must belong to one preserved execution contract.
Agent runtimes
Hermes packages three days of rapid change into a stable tag
Hermes Agent 0.20.2 combined, by the maintainer's count, about 397 pull requests, 967 commits, and 1,279 files between the August 13 and 16 tags. The rollup spans multiple agent gateways, MCP health, persisted model routes, prompt caching, Windows probes, Telegram topics, and hardening of cron, authentication, and installers.
The counts include generated and mechanical changes and do not equal features. Full curated notes were deferred to 0.21.0, so the release cannot support claims more specific than the maintainer provided.
The stable checkpoint itself matters. A fast-moving runtime needs a reproducible version for downstream Docker images, hosted deployments, and fresh installs. Without a tag, "latest main" is not a shippable product and cannot tie a failure to a particular build.
Source: Hermes Agent 0.20.2.
Pricing
DeepSeek's rate changes in the middle of the day
The V4-Pro peak/off-peak schedule announced on August 13 took effect on the 16th at 16:00 UTC. Its table distinguishes cache hits and misses, input and output, and time of day; the model itself also offers low, high, and max reasoning.
Without a concrete token mix and local time window, the rate cannot be reduced to one price or percentage change. An official schedule is not a measured customer bill.
The exact timestamp is essential in retrospect: a morning and evening run on the same day could use different rules. An agent ledger must store not only the model name but pricing version, cache behavior, and effort for every call. Otherwise historical cost cannot be reconstructed honestly.
Source: DeepSeek pricing schedule.
Autonomous research
Eight days of work did not guarantee an agent new scientific ideas
Prime Intellect ran 153 autonomous research trials across 18 frontier models. Each agent received an isolated 8xH200 environment and up to eight days to reduce the training steps needed for a 124M GPT to reach target loss. The best Claude Fable 5 run reached 2,726 steps—81.7% of the path from the 3,290 baseline to the claimed human record of 2,600.
That is a share of the gap between selected points, not a "percentage of human intelligence." The date comes from a contemporaneous announcement, compute budgets were enormous, and the human record was not independently verified.
The authors report high variance and few genuinely new ideas. A long horizon allows more variants to be tried but does not automatically create research taste. We need replicates, a hypothesis log, and separation of a discovered improvement from one lucky run.
Source: Measuring Autonomous Research.
Memory architecture
HyMem keeps detailed execution from flooding the high-level plan
HyMem stores the plan separately from execution traces. Complex analysis moves into an isolated module and returns a structured summary to persistent context. With DeepSeek-V4, the authors claim Pass@1 of 66.7% on GAIA and 61.3% on BrowseComp-plus—6.1 and 4.7 points above the strongest baseline.
This is two benchmarks and one backbone. Summary cost and the risk of losing evidence are excluded from the headline; there is no independent replication.
The separation is useful as more than another compression technique. The plan answers where the task is going; the trace, what actually happened. Mix both into one growing context and a late detail displaces the goal. A summary should link back to its source trace so that savings do not destroy verifiability.
Source: HyMem.
Delegated authority
Authority narrows at every handoff between user, agent, and subagent
Agentic Principal Chain stores scope and delegation budget outside the model and checks each new action in the context of prior grants. Across 3,154 instances, its compromised-model test reduced AgentDojo exfiltration from 75–100% to zero and blocked 544 data-theft cases. Intent binding reduced destruction from 38.6% to 4.0% and manipulation from 90.5% to 12.1%; p99 verification was 0.24 ms.
The attack call was inserted from ground truth after the first legitimate call, so the experiment isolates authorization rather than discovery. Utility fell by 8.6 and 13.9 points in two settings. An unknown dangerous composition remains allowed if the restriction set does not describe it.
The result reveals the price of real blast-radius control. Authority is not copied wholesale to a subagent; it becomes narrower, while verification remains stateful and serialized. Without this chain, delegation silently becomes escalation.
Source: Bounded Agents.
Review
Three agents became useful only after disagreement was mandatory
Adversarial Review uses a coding agent, reviewer, and critic. The critic must test the review through structured disagreement before fixes. On LiveCodeBench, three agents produced the best pass rate among the methods and beat a five-agent baseline. On SWE-PRBench, a naive design produced false consensus; one prompt iteration requiring explicit disagreement yielded the best F1.
The abstract gives no absolute scores, and the prompt was changed after the problem appeared, possibly fitting the benchmark. This is not a production code review study.
But the mechanism fits the collaboration tax. More agents are not useful by themselves. If all seek quick agreement, more votes merely amplify one error. The critic role must require independent rechecking, not stylistic approval of the reviewer.
Source: Adversarial Review.
GUI agents
One successful demonstration doubles strict success—on the same targets
UI-Mate-27B combines a grounded data engine, SFT, online RL, and GUI demonstrations. OSWorkerBench contains 100 long office tasks across 41 applications. The authors claim 77.0% on OSWorld-Verified, 66.2% on WindowsAgentArena, and 41.0% strict success on their own benchmark.
On 33 self-demo tasks, one successful strong-agent trajectory raised strict success from 17.2% to 35.4%. But these are demonstrations of the same targets; separate variant tasks use related human recordings. The same team owns the new benchmark and model.
A demonstration is powerful because it shows not just the goal but a path through a specific UI. That same path can easily overstate transfer. Evaluation must separate instruction-only, same-target demonstration, and related-task demonstration.
Sources: UI-Mate paper, project.
Serving efficiency
A small full-context helper rescues a large model after aggressive eviction
KV-Rescue alternates the main model's reasoning with a full-context helper and stops degenerative candidates using entropy and compressibility. Across five math benchmarks with Qwen2.5-Math 7B and 72B at cache budget B=64, the method recovered an average 87% of lost accuracy; early stopping cut base-model generation by 43%.
The 87% is measured against the gap between evicted and full-KV performance, not as absolute accuracy. Fewer large-model tokens do not equal lower total cost: the helper also consumes compute and context. Only the math/Qwen setup was tested.
The complementary error is what matters. A small model with complete memory can correct a large one that lost necessary context. Backbone size and access to evidence are independent resources; more parameters do not compensate for a deleted fact.
Source: KV-Rescue.
The issue's main technological shift
On August 16, the long task acquired its own execution contract.
A tag fixes the runtime. A pricing timestamp fixes cost. The plan is separated from the trace. Authority narrows through delegation. The reviewer must argue. A demonstration is labeled by proximity, and a helper restores lost evidence.
An agent can work for eight days only if the system can reconstruct which version, authority, price, and evidence governed every step.
What to discuss with the technical team
- Can every failure be tied to an exact runtime tag and configuration?
- Is the pricing version of every model call retained for retrospective economics?
- Does a summary link back to its full execution trace and evidence?
- Does authority narrow at every subagent handoff?
- Must an independent reviewer provide evidence-backed disagreement rather than a second vote?