One agent prepares commercial proposals and tender documents, another deploys clusters and investigates failures, a third finds specialists. Each has its own area, its own tools and a way to check the result.
A working route: from an application error to a fix
idevopsmonitoring sees errors rising
alert-relaythe signal becomes a task
GitHub issuein the product repository; a repeat updates the open one
iompreceives the event from GitHub
roboomp · ompa development agent looks for the cause
pull requesta comment with findings or a fix
Raw logs never go into the task: they may hold personal data and tokens.
01Agents
Each agent has its own job
iomp runs development tasks but does not replace specialisation: ipresale knows tenders, idevops knows infrastructure, italent knows hiring. Each needs its own instructions, data and access.
ipresale
Commercial proposals and tenders
passport
fetch
extract
pack
documents
price
pricelist
forms
kp
contract
assemble
submit
haggling
errors
handoff
Collects requirements from four sources, prices the offer, fills in the platform forms and assembles the submission package.
ResultSubmission package, inventory, field map and a list of open questions
Takes an issue, starts a development agent on roboomp and omp, returns a root-cause analysis or a pull request.
ResultA comment with findings or a pull request linked to the task
02Anatomy
What an agent is made of
The model chooses actions and reads the content. Repeatable operations are done by tools: they download files, add up sums, fill in cells and run checks.
The order of steps, result requirements and stop conditions live in the repository. A process fix carries over to the next task.
ipresale15 stages, each with an input, a result and a check
idevopslocal runbook: host, cluster and the limits of change
italentthe score is computed by code from the checklist, not asked of the model
03Limits
Where an agent stops
Permissions are set per agent and per environment. Preparing something does not grant the right to an action with consequences.
The agent does itA person decides
Prepares the submission package and field mapSubmit the bid: a submission is an offer
Opens a pull request with a fixRelease the change
Reads code, logs and cluster stateDelete data or widen the scope of change
Shows a gap against a criterionBid with an incomplete match
Unfinished work is flagged explicitly: missing data, a failed check, no access or a decision needed. A model saying "done" does not replace that.
04Measuring
How to measure the effect
By finished work, including checks and corrections. The number of agents, files or model calls describes the setup, not the savings.
01Preparing a proposalTime to a checked package, manual edits, missed requirements, open documents
02DevelopmentTime to an accepted change, criteria passed, regressions and reopenings
03OperationsTime to diagnosis and recovery, repeat incidents, wrong actions
04Task handoverLost requests, duplicates, waiting time, tasks with no owner
05Start
Tell us which work to hand to an agent
We will go through the process, the limits and how to check the result.