Cohesor's bet: the agent gateway becomes the new cloud cost layer

Cohesor is launching on Product Hunt as a “neutral control plane for enterprise AI agents,” a middleware layer that sits between agents like Claude Code or Cursor and the underlying LLMs. The pitch is aggressive: cut agent model spend by 60–90%, compress input tokens by roughly 50% on average, and add only about 12 milliseconds of latency at p50. The company’s own site frames the mission in stark terms: agents now “write, test and ship real code” by “sending enormous prompts to the most expensive models available, thousands of times a day,” creating “a new category of spend that scales with your team’s productivity, with almost no tooling to understand or control it.”

The genuinely new thing here is not compression or routing. Those techniques have existed in various forms since the early days of LLM API gateways. What is new is the position Cohesor claims: a control plane that is “neutral by design, owned by no model vendor,” explicitly refusing to sell a model or train on customer data. That positioning is a direct response to a market where the biggest players have obvious conflicts. Anthropic, OpenAI, and Google all sell both models and, increasingly, agent tooling. A gateway that routes to “the best model for each request” is structurally at odds with a vendor that profits when you stay on its models. Cohesor is betting that enterprises will pay a premium for a layer that has no incentive to steer traffic.

The economics of agent fleets make the timing credible. As Cohesor notes, agents send “enormous prompts” thousands of times a day, and the cost driver is overwhelmingly input tokens: file contents, conversation history, tool outputs, all re-sent on every turn. The company’s compression engine targets exactly that. Its techniques include stale-read supersession, which removes outdated file content; context compaction, which folds out stale conversation turns; and semantic deduplication, which collapses near-identical text spans above a 0.97 similarity threshold. These are not exotic research ideas. They are engineering solutions to a problem that every serious agent team has hit: the context window is a billing meter, and agents fill it constantly.

The more interesting claim is the routing layer. Cohesor uses a “lightweight complexity classifier” to score each prompt and route it to a model based on configurable objectives: Quality, Balanced, Cheapest, or Fastest. The example in its materials is telling: expensive models like Claude Opus are reserved for complex tasks, while simpler queries go to cost-effective models like Llama 70B on Groq. That is a meaningful architectural statement. It implies that a meaningful fraction of agent traffic does not need frontier models at all, and that the bottleneck to cheaper agent operations is not model capability but routing intelligence. If that holds, the agent cost curve flattens faster than the model providers would like.

The governance story matters just as much. Cohesor brokers MCP (Model Context Protocol) tools through a single audited endpoint, managing OAuth 2.1 credentials, enforcing per-agent tool scopes, and logging every tool call with full traces. The dashboard attributes spend per user, API key, and team, with hard budget caps enforced at the network edge. This is the part that finance teams will care about. The company’s own quote captures the ambition: “We want running an agent fleet to feel as governed and observable as running a cloud account.” That is the right comparison. Cloud cost management became a category precisely because engineers adopted cloud services faster than finance could track them. Agents are repeating that pattern, only faster.

The skeptical read is that Cohesor is early and unproven. The company is hiring founding engineers and a developer advocate, which suggests a small team. Its headline numbers, 60–90% spend reduction and ~50% token compression, are presented without independent verification. The 12ms p50 latency figure is a self-reported benchmark. The “zero code changes” claim, while technically plausible via a one-line environment variable change like ANTHROPIC_BASE_URL, assumes that the compression and routing heuristics hold up across the messy variety of real agent workloads. Lossless compression is a strong claim. Any semantic degradation, however rare, would surface as subtle agent misbehavior, and that is the kind of failure that erodes trust in a governance layer.

Still, the strategic position is the story. Cohesor is making a bet that the agent era will produce a new infrastructure layer, and that this layer must be neutral to be trusted. That bet has a clear precedent: the cloud itself. AWS, Azure, and GCP built multi-tenant infrastructure that abstracted away the underlying hardware, and a whole ecosystem of observability and cost tools grew on top. The agent gateway is the analogous layer for the model era, and the question of who owns it is already contested. Cloudflare has pushed AI gateways. LiteLLM and Portkey offer routing layers. The model vendors themselves are building agent platforms. Cohesor’s differentiation is the neutrality guarantee, and that guarantee is only as strong as the company’s independence.

The tension is that the most valuable position in any infrastructure layer eventually attracts the giants. If Cohesor succeeds, it will face pressure from the model vendors it routes around, the cloud providers it sits on top of, and the agent frameworks it brokers. Its answer, that it never trains on customer data and sells no model, is a durable moat only if enterprises actually value neutrality enough to pay for it. The early evidence suggests they might. Agent spend is opaque, unbounded, and increasingly material to cloud bills. A control plane that makes that spend visible, governable, and cheaper is solving a problem that finance teams are already feeling.

For AI builders, the implication is direct: the cost of agents is not fixed by model prices. It is a function of how you route, compress, and govern. Cohesor’s numbers, even discounted for marketing, point to a reality where a large share of agent traffic is overpaying for frontier models and re-sending redundant context. The tools to fix that are emerging now, and they are coming from a layer that did not exist two years ago. The question is not whether agent cost optimization becomes a category. It already has. The question is whether the category winner stays neutral long enough to matter, or whether the model vendors absorb the control plane into their own platforms. Cohesor is placing its bet on the former. The next year of agent adoption will test whether that bet holds.