The Agent Has Left the Chat Window

For nine years, the transformer has been both the industry’s engine and its zoning law. It made modern language models possible; it also made every ambitious product roadmap quietly dependent on ever more compute, memory and electricity. Dense attention compares tokens with one another at punishing scale: the MIT Technology Review account estimates that a 10,000-word document can entail roughly 50 million multiplications. Context windows have become less a feature than a municipal infrastructure project.

That creates a useful corrective to the usual model-release liturgy. A group of startups is trying sparse attention, rolling context summaries, liquid neural networks, diffusion generation and state-space methods—not because transformers suddenly became unintelligent, but because intelligence with an alarming utility bill is a constrained resource. The reporting also supplies the important asterisk: many of the impressive comparisons are company claims without disclosed benchmark methodology, hardware settings or independent replication. The demo is not the due-diligence packet.

Meanwhile, Meta’s Muse Glimmer announcement offers the incumbent version of the same thesis: a 30-billion-parameter, Apache-2.0 open-weights model aimed at local agent workflows. Meta says 4-bit quantization brings its language model below 20 GB, leaving room on 24 GB or 32 GB systems for cache, perception and a speculative-decoding drafter. The industry’s real contest is therefore not simply bigger versus smaller. It is whether useful capability can be packed into a deployment envelope that does not require a purchasing department, a data centre and a modest weather system.

Local models are becoming plausible precisely as the permission surface gets less plausible. A coding agent that merely proposes a patch is a fancy autocomplete. A coding agent that installs packages, alters configuration, runs services, calls tools and continues while its operator sleeps is an employee with root-adjacent hobbies. The technical achievement is autonomy; the operational question is where, exactly, to put it.

Docker Sandboxes is an unusually candid product response. Docker markets disposable local microVM environments for agents including Claude Code, Gemini CLI, Copilot CLI, Codex and others, with the project workspace mounted in and configurable network and filesystem controls. It explicitly labels --dangerously-skip-permissions as “YOLO mode.” This is, for once, accurate product nomenclature: speed comes from removing the human checkpoint, while risk comes from removing the human checkpoint. Product management has solved a difficult naming problem by declining to lie.

Docker’s pitch is marketing, not an independently published security evaluation, so treat the promised boundary as something to verify against your threat model rather than a sacrament. But the direction is sound. As agentic systems gain the ability to operate tools and persist through long tasks, containment becomes part of the application architecture, not a compliance appendix. Local inference shifts where data and computation reside; sandboxing determines what the model may do when it is confidently wrong. Those are separate controls. Teams that collapse them into one word—private, secure, local—are preparing an incident report with excellent branding.

The common thread is not open versus closed, nor transformer versus whatever comes next. It is the unglamorous engineering of constraints. Sparse attention is an attempt to spend fewer operations. Context retention is an attempt to remember selectively. Diffusion-style text generation and speculative drafting seek more output per unit time. Quantization is an attempt to make a large model fit where actual developers work. Sandboxes are an attempt to stop that conveniently placed model from redecorating the workstation.

This matters for CTOs because the architecture changes the economics before it changes the philosophy. Liquid AI says some of its models can run on a $50 Raspberry Pi; Inception says its diffusion approach can generate text substantially faster; Pathway says its state-space system can handle non-sequential reasoning tasks such as Sudoku. These are intriguing claims, reported with meaningful caveats, not permission to rewrite the platform roadmap over lunch. Ask for reproducible evaluations, target-hardware throughput, quality under your workload, failure modes and total operating cost. A benchmark score without a deployment profile is a résumé written by the candidate’s publicist.

The emerging stack may be less theatrical than the frontier-model race: a right-sized model, compressed enough for the intended machine, accelerated where latency matters, granted only the permissions its job requires, and disposable when it misbehaves. Meta’s release promises local agent operation; Docker’s product promises a place to let such agents act with limits. Put those together and the next wave of AI may look less like an oracle in the cloud than a junior operator in a locked workshop. The lock is not a concession to the future. It is the part that makes the future deployable.