I shipped adaptive mode for the Ejentum reasoning harness. Here's what changed and why it matters if you build agents.
The harness, in one paragraph
Your agent calls a tool mid-task and gets back an engineered cognitive operation: the failure to avoid, a step-by-step procedure, a reasoning topology (a DAG of steps and decision gates), suppression signals, a falsification test, and a target pattern. The agent reads that before it answers. It is not RAG (that fixes what the model knows) and not chain-of-thought (that just makes reasoning visible). It governs which reasoning steps to take and which failure modes to block.
What was there: dynamic
Dynamic delivery does a single retrieval. The highest-scoring operation comes back as engineered. It is predictable and fast, but the operation is canonical, written against a generic example, not against the problem in front of your agent.
What's new: adaptive
Adaptive does top-k retrieval, then an adapter LLM rewrites two fields, the procedure steps and the reasoning topology, with the identifiers from your actual task. The agent gets a cognitive map fitted to its exact problem instead of a template.
The part I care about most: the safety-critical fields (negative gate, suppression signals, falsification test) are withheld from the adapter and stitched back in by code after it runs. The adapter only ever sees the task, the procedure, and the topology. So adaptive can fit the reasoning to your task but cannot weaken the guardrails. That property is structural, not a promise.
Live
The screenshot is a Claude instance calling adaptive-reasoning mid-task. The returned topology referenced the actual work it was doing (S1:decompose(ejentum_portrait -> components ...), cognitive style recursive self model), not a canonical example. That is the whole point: the map matched the territory.
Use it
REST:
curl -X POST "https://api.ejentum.com/harness/" \
-H "Authorization: Bearer YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"query": "diagnose why a microservice returns 503s under load", "mode": "adaptive-reasoning"}'
MCP: point any MCP client at https://api.ejentum.com/mcp and the tool harness_reasoning_adaptive appears (plus dynamic and the other three harnesses).
Honest notes
Adaptive adds one adapter round-trip, so it is a couple seconds slower than dynamic. Use it when the task is novel enough that a generic operation would not fit.
It draws on a separate adaptive call pool. The 30-day free trial is dynamic-only.
If the API is unreachable, your agent continues on native reasoning. It is an enhancement layer, not a dependency.
Same one tool call. From an abstract, fixed operation to an adaptive cognitive map, on the fly.
Harness the full reasoning power of your agents. → https://api.ejentum.com/mcp
Repo: https://github.com/ejentum/ejentum-mcp
Quickstart: https://ejentum.com/docs/quickstart
Top comments (0)