Product
INFRASTRUCTURE FOR AGENT HEALTH.
Elixir sits beside your AI workforce and answers one question well: this agent says something is wrong — is it, and what should be done about it? It receives the incident, identifies the durable protected agent, observes approved runtime state, collects evidence, diagnoses with a provider you configured, attempts only authorized remediation, refuses unnecessary work, verifies the result independently, and returns a canonical recovery result.
An Elixir Agent ID is issued once and never reused. One logical agent keeps one identity — and one history — through changes of host, runtime, worker, provider, repository and name.
Runtime adapters read service state, health and repository state. None of them can start, stop or restart anything: there is no write path in an observation adapter.
A repair runs only where a capability a human wrote down permits it, inside limits on turns, tool calls, transcript size, wall clock and daily spend.
If the evidence does not support treatment, Elixir changes nothing and records why. That is a completed run, and one of the two cases every certified executor must pass.
The session does not grade itself. Elixir re-runs what the session claims passed, judged by an oracle the provider did not control.
Two proven executors, eight more implemented. The provider supplies reasoning; Elixir supplies the recovery system, and never silently swaps one provider for another.
An incident, end to end
WHAT HAPPENS AFTER AN AGENT ASKS FOR HELP
AI agent
│ reports an incident
▼
Elixir MCP ─or─ Public incident API
│ authenticated; holds no treatment authority
▼
Identify the durable protected agent
│ agt_… — issued, never derived
▼
Observe approved runtime state
│ read-only
▼
Collect evidence, with provenance
▼
Diagnose (your configured provider)
▼
Repair ─or─ Refuse
│ bounded by written capability
▼
Verify independently
│ the provider does not grade itself
▼
Canonical recovery resultIncident states
- queued
- Accepted, not yet worked
- diagnosing
- A session is running
- needs approval
- A human gate is open
- completed
- Verification accepted it
- refused
- No repair required
- blocked
- Cannot proceed; a person must look
Who it is for
TEAMS RUNNING AGENTS THAT HAVE TO KEEP RUNNING
Elixir earns its place when an agent is persistent, does real work, and failing quietly is expensive.
Coding agents
Long-running agents on real repositories, where a broken dependency or a drifted config stops work until someone notices.
Operations fleets
Supervised processes across hosts, where "the service says it is healthy" and "the service is healthy" are not the same claim.
Customer-facing agents
Where a silent failure is not an inconvenience but a queue of unanswered people.
Research and data agents
Long pipelines where a failure at hour two is discovered at hour nine.
Scope
WHAT ELIXIR IS NOT
Not an agent framework
It does not build or run your agents. It recovers the ones you already have.
Not a model provider
It has no model of its own and resells no tokens. You bring the provider.
Not an APM or a log aggregator
It observes what a runtime adapter is permitted to read, for the purpose of diagnosis — it is not a metrics pipeline.
Not a hosted service
Elixir is self-hosted. There is no control plane, and no credential of yours reaches AgentDoctorMD.
Not an approval authority for itself
It cannot grant itself a permission, and neither can any interface it exposes.