The agent stack beyond the model: sandboxes, harnesses and memory
Five releases this week aim at the same problem: making agents cheap, fast and correct to run, not just capable.
- Daytona says it provisions Firecracker microVMs in less than 90 milliseconds, while agent workloads on Kubernetes need custom controllers for lifecycle management and provisioning. Self-reported m:…4999
- Factory says Droid 1.0 outperforms native model harnesses by up to 20% in efficiency on the same tasks, with Programmatic Tool Calling running hundreds of tool calls in a single model request. Self-reported m:…3345
- Baseten says joey-1, a 2-billion-parameter decision model, has a P50 end-to-end latency of 20 milliseconds, costs $0.01 per million tokens, and is 2.5 times faster than elephant-1 on the same latency measure. Self-reported m:…7269
- LlamaIndex says max effort for Extract raised Value F1 on ExtractBench Pro from 77.7 with standard Agentic Plus to 92.9, with variable billing based on the computation required. Self-reported m:…0477
- Zep says its Context Graphs keep two timelines for every fact, so a contradicted fact is kept and given an end date rather than deleted. Self-reported m:…1655
Why Kubernetes alone falls short
Daytona's argument starts with a mismatch. Kubernetes was designed for long-running microservices where instances are interchangeable, but AI agents require short-lived, task-specific environments that adapt in real time. Running agents on Kubernetes, Daytona says, forces teams to build custom controllers for lifecycle management and provisioning. m:…4999
Daytona's answer is to provision Firecracker microVMs in less than 90 milliseconds, and it points to customers to make the case. AfterQuery reduced sandbox creation times by 86%, Trajectory saves more than 15 hours a month by using Daytona instead of stitching together Kubernetes solutions, and Athena Intelligence consolidated six specialized sandbox environments into a single runtime layer. WolfBench schedules sandboxes across cloud infrastructure, virtual machines and bare-metal servers to minimize idle compute costs, and the Laude Institute increased experiment throughput 100-fold using Daytona's stateful environments, according to Daytona. m:…4999
The harness sets the efficiency
Factory's Droid 1.0 takes the same question one layer up. The company describes it as an agent harness designed for enterprise development that works independently of any single model, supporting both open-weight and closed-weight frontier models. Its default tool-calling mechanism, Programmatic Tool Calling, lets agents execute hundreds of tool calls in a single model request. m:…3345
Factory says Droid 1.0 outperforms native model harnesses by up to 20% in efficiency on the same tasks. It also reports a 70% reduction in median time to first paint and time to accept a prompt compared with August 2026, and a 37% smaller base prompt between versions 0.170.0 and 0.250.0. A daemon-based architecture shares memory across sessions, cutting memory use per session to 20 MB, and the harness runs on macOS and Windows desktops. Factory also runs an internal loop, called Agents Factory, that evaluates new models against benchmark failures and real-world user friction. m:…3345
Cheap decisions, pooled capacity
Not every agent decision needs a frontier model. Baseten's joey-1 is a 2-billion-parameter decision model that returns a label and a confidence score in one forward pass, with a P50 end-to-end latency of 20 milliseconds and a price of $0.01 per million tokens. Baseten says it is 2.5 times faster than its elephant-1 model on the same latency measure, and it offers joey-1 with open weights, a 16,000-token context window on its Model APIs (262,000 tokens on dedicated deployments), and zero data retention by default. m:…7269
Baseten is also the recommended inference partner for Postman's enterprise agents. Postman's Agent Stack includes Passport, Fabric, Astropods and Fern, and Baseten pools GPU capacity across clouds with its Multi-cloud Capacity Management, while Baseten Training provides infrastructure to post-train open models for specific tasks. Postman's platform is used by more than 40 million developers and 98% of the Fortune 500, according to Baseten, and its model APIs ranked among the top two providers on Artificial Analysis for throughput and latency in 2026. m:…4786
Extraction quality and memory correctness
LlamaIndex's Extract V2.6 adds a max effort setting to its Extract agentic_plus tier, aimed at critical and complex documents. On ExtractBench Pro, Value F1 rose from 77.7 with standard Agentic Plus to 92.9 with max effort, and the setting can examine relevant regions at higher resolution to read visual charts better. The trade is explicit: max effort uses variable billing based on the computation required and can cost more than a standard job, while standard tiers absorb variability within a fixed price per page. LlamaIndex recommends running a representative sample of documents with existing schemas before migrating, and notes its max effort benchmarks were measured on October 29, 2026, against standard benchmarks run October 9 to 14. m:…0477
Zep's contribution is about memory rather than speed. Its Context Graphs keep two timelines for every fact: an event timeline recording when a fact was true in the world, and a system timeline recording when the graph learned the fact and when it learned the fact ended. When a new fact contradicts an old one, the old fact is kept and given an end date, so the graph preserves what was true and when it stopped being true. A reference time marks when the source event happened, which lets the system resolve relative dates like last week, and the order in which facts arrive does not change their validity dates. That detail matters because backfills can deliver old facts after the facts that replaced them. Most internal data agents, Zep says, check for stale context only by testing how recent it is. m:…1655
What to weigh
The through-line is that the expensive part of an agent system is shifting from the model call itself to everything around it: provisioning the environment, batching the tool calls, remembering what changed, and deciding when a cheap model or a harder extraction pass is enough. That makes vendor-reported efficiency numbers the thing to scrutinize, since every figure in these releases comes from the vendor itself. A team adopting any of these pieces should ask for the same task run on its own workload before trusting the headline percentage. m:…4999 m:…3345 m:…0477
What is still unresolved
- All efficiency, latency and benchmark figures in this story are vendor-reported; none has been confirmed by an independent benchmark in the material reviewed.
- None of the releases reviewed describes how these five pieces interoperate, so integration effort between them is unknown.
- What max effort costs on a representative workload is not stated beyond that it can cost more than a standard job.
- Whether two-timeline memory measurably improves agent accuracy is not measured in the material reviewed.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.