Service

Consistent incident diagnosis on every shift, with people approving every action beyond agreed limits

We build an AI agent into your incident process, working on your runbooks, telemetry and device state. It gathers the context an engineer would, analyses it and proposes the most likely cause with its evidence. Actions come in agreed scopes, and a person approves every step outside them.

Measured impact

What changes for your organization

  • Today

    Engineers buried in alerts

    When something breaks, syslog and alerts multiply faster than the on-call engineer can read them.

    With assisted diagnosis

    One incident, one picture

    Repeated events are grouped and condensed into one alert with its severity, likely causes and next steps.

  • Today

    Diagnosis that depends on the shift

    How fast a fault is found depends on who is on call and how well they know that part of the network.

    With assisted diagnosis

    The same investigation every time

    Every incident is investigated the same way, on your runbooks and past tickets, whoever is on shift.

  • Today

    An investigation across many tools

    Topology, device CLI, logs and alerts each live in their own tool, and the engineer stitches them together by hand.

    With assisted diagnosis

    Evidence gathered for you

    The agent pulls config, state, logs, alerts and topology together, so engineers start from the evidence.

  • Today

    Senior engineers on every escalation

    Hard faults go to the same few experts, and their time sets the pace of every incident.

    With assisted diagnosis

    Faster time to cause

    Ranked hypotheses with their evidence reach the first responder, and experts step in where judgement is needed.

  • Today

    Automation too risky to switch on

    An agent that changes production unchecked is a risk the organization will rightly refuse.

    With assisted diagnosis

    Autonomy on your terms

    Every proposed change is verified before it applies, and a person approves each action outside the limits you set.

What we deliver

  • 1. Detect and condense

    Errors counted per device; once they cross a threshold, an alert opens. Repeated events are grouped by device, process and error code, and similar messages are merged into one.

    Read the case study →
  • 2. Separate concurrent faults

    Messages clustered by meaning into issue buckets, so two faults at once, or problems across protocols and layers, are analysed one by one.

  • 3. Gather the evidence

    The agent plans what else it needs and fetches it through MCP tools: topology, interface and BGP or BFD state, live config, alerts, and past tickets with similar symptoms.

    Read the case study →
  • 4. Rank the likely causes

    The engineer gets the alert with several likely causes, each with a probability, the evidence behind it and the suggested next steps from your runbooks.

  • 5. Propose and verify the fix

    Fixes ranked from surgical to blunt with their blast radius, and each checked in a Verify Gate: a change-stop, a dry run on live state and a prepared rollback.

    Read the case study →
  • 6. Approve, apply, record

    A person approves each action outside the agreed limits, the result is checked on the device, and every step lands in an audit journal.

How it runs

Every engagement starts with assisted diagnostics on your own incidents; each wider scope is agreed once the first proves itself.

  1. Discovery

    Your incident process, runbooks, telemetry and past tickets, the incident types to start with, and a plan with the scopes and limits.

  2. Data access

    Syslog, metrics, alerts, topology and device state connected to the agent through MCP tools, with role-based access and an audit trail.

  3. Diagnosis on past incidents

    The agent replays incidents you have already solved, and its causes are checked against what your engineers found.

  4. On-call rollout

    Assisted diagnostics in your live incident process, beginning with the incident types your team sees most.

  5. Wider scope

    Guided remediation or bounded autonomy, when agreed, with the Verify Gate and the limits set per class of action.

  6. Handover

    Your team runs and extends the agent, its tools and its runbooks, documented and owned in-house, with our support as long as you want it.

Three scopes, each agreed separately

Assisted diagnostics is the standard engagement. Each next scope is agreed on its own, once the one before has earned trust.

  • Assisted diagnostics

    The agent gathers context, analyses the data and proposes the cause. The standard engagement.

  • Guided remediation

    The agent proposes actions, and a person approves each one before it applies.

  • Bounded autonomy

    The agent carries out approved classes of action alone, within set limits. Everything beyond them goes to a person.

Proof

Our work here comes from R&D prototypes on live lab networks and from production telemetry pipelines.

Why CodiLime

  • Network engineers who build agents

    The team behind Net-Inspector knows BGP, fabrics and syslog, and built the agent that diagnoses faults on a ~200-device lab network.

  • Control built in from the first step

    Our bounded-autonomy prototype passes every change through a Verify Gate, sends larger ones to a person and records every step in a hash-chained journal.

  • Open standards, your choice of model

    Tools exposed through MCP stay the same under several LLM backends, as on Net-Inspector, so you can change models as they improve.

FAQ

Good to know

  • Will the agent change our network on its own?

    Only in the scope you agree. The standard engagement is assisted diagnostics: the agent proposes the cause and people act. Guided remediation and bounded autonomy are agreed separately, with the limits set per class of action.

  • What data does the agent need?

    What your engineers use in an investigation: syslog, metrics, alerts, topology, device config and state, runbooks and past tickets. The discovery shows which of them you have and how to connect them.

  • Which AI model do you use?

    The one that fits your rules on data and cost. The tools sit behind MCP, so the same tool layer works under several models, hosted or run in your environment, and a model can be replaced later.

  • How do we know the diagnosis is right?

    Before it reaches on-call, the agent replays incidents your team has already solved, and its causes are compared with theirs. In production every hypothesis comes with the evidence behind it, so the engineer can check it.

  • How is the agent itself kept under control?

    Role-based tool access, a policy gateway, a Verify Gate before any change and an audit journal of every step. An action outside its agreed estate is refused.

Start with an incident diagnosis assessment

Tell us how incidents run today. We look at your alerts, telemetry, runbooks and past tickets, pick the incident types to start with, and hand you a plan with the scopes and limits for the agent.