Open sourceSelf-host it or use our cloud

Build a software factory you actually own.

Define your agents’ workflow as code, run it on any model and measure every change.

Coming soon

The Delivery line in Spotlight: tasks in Triage, Spec, Implementation and Code Review, two of them waiting for a person

lines/delivery.tf

1resource "spotlight_harness_profile" "careful" {
2  team_id = spotlight_team.platform.id
3  key = "careful"
4  permissions = {
5    allow = ["read", "edit", "execute(npm test *)"]
6    ask = ["execute(git push *)"]
7    deny = ["read(.env)"]
8  }
9}
10
11resource "spotlight_line" "delivery" {
12  team_id = spotlight_team.platform.id
13  key = "delivery"
14  name = "Delivery"
15
16›  station { ··· } # Triage
22›  station { ··· } # Spec
28
29  station {
30    key = "implementation"
31    name = "Implementation"
32    harness = "codex"
33    model = "gpt-5"
34    profile = spotlight_harness_profile.careful.key
35    tools = ["linear.comment", "github.comment"]
36    skills = ["grafana"]
37    check = "npm test"
38  }
39
40›  station { ··· } # Code Review
46}

From an idea to production, and back.Spotlight runs every stage of the lifecycle.

Agents write code in hours. Specifying, verifying and operating it is now the bottleneck.

1. Plan
Turn a question into planned work
Agents ship a clear ticket in hours. Specification is now the bottleneck.
Your team’s agent scopes the work and proposes tasks. Nothing starts until you accept.
2. Spec
Agents ask before they guess
Agents rarely fail on syntax. They fail on wrong assumptions.
The Spec station writes the spec first and asks when the decision is yours.
3. Build
Every station runs its own agent
One general-purpose agent is hard to steer and spends frontier tokens on routine work.
Each station runs its own agent, prompt, tools and model.
4. Review
Review the change and the reasons
Review is the new bottleneck: more pull requests, larger diffs.
Spotlight groups each change into reviewable layers, each with the agent’s reasoning.
5. Operate
A first responder for every alert
Shipping faster means more alerts for whoever is on call.
Alerts become tasks. The agent investigates and asks before it touches production.
6. Learn
Every incident improves your observability
Generic agents don’t know your dashboards, alerts or conventions, so the same incident happens twice.
Spotlight knows how your team works and the tools it uses. After a fix, it proposes the missing alert, Grafana panel or memory.

You can’t own what you can’t measure.Every change to your harness, measured.

Most agent failures come from the harness, not the model. Improving it requires measurement.

7. Improve

Config changes in Spotlight Insights: four changes to the Delivery line, each with runs before and after and an Impact button

The impact of a change at Implementation: effort raised and a skill added, with sent-back rate up 18 points and cost per run up 61%

See what every change did

Prompts and models change weekly. Without numbers, every change is a guess. Every config change is tracked against runs, rework and cost.

A replay of a shorter prompt on billing work: 60% success, 20 points below the baseline, task by task

Test a change before it ships

Testing on live work makes your team pay for regressions. Replay a change on past tasks. Keep it only if it wins.

The harness is your team’s asset.Keep it on your terms.

Labs and vendors want to own how you build software. Spotlight keeps it yours.

The Runners page in Spotlight: build machines and a Mac mini in a pool, each running tasks

Runs on your machines

Agents need your code, credentials and systems. Run them on your servers, in sandboxes like Daytona or E2B, or in our cloud. Use a built-in runner or add your own through an integration.

Integrations in Spotlight: GitHub, Linear, Sentry, Slack, Grafana and PagerDuty, next to the team’s own Payments API, Billing API, Honeycomb and Statuspage

Your integrations, your tools

Every team relies on internal APIs no vendor supports. Connect your tools, or add your own in a few lines.

The harness menu in an agent chat: Claude Code, Codex and Cursor, with model and effort settings

Any harness, any model

The best model changes every few months. Choose the harness and model per station.

integrations/billing.tf

resource "spotlight_integration" "billing" {  team_id = spotlight_team.platform.id  definition = {    integration = "billing-api"    actions = {      refund = {        request = {          method = "POST"          url = "https://billing.acme.internal/refunds"          body = { invoice = "$${{ inputs.invoice }}" }        }      }    }    tools = {      refund = {        does = "refund"        description = "Refund an invoice."        inputs = { invoice = { type = "string" } }      }    }  }}

Configured as code

An unmanaged harness drifts. Everything is Terraform, reviewed in pull requests.

Skills in Spotlight: grafana-dashboards, our-migrations and release-notes, each with what it teaches

Skills your team writes

Agents don’t know how your team does things. Write it down once as a skill, and give it to any station.

Connect the tools your team uses,or define your own integration.

Review pull requests and fix failing checks
Pick up issues and report back on them
Answer an agent’s question in the thread
Investigate alerts as they fire
Open work from incidents
Turn errors into tasks
Update status during incidents
Add your own API in a few lines

Agents write the code.You own the factory.

Open source. Self-host it or use our cloud.

Coming soon

Build a software factory you actually own.

Turn a question into planned work

Agents ask before they guess

Every station runs its own agent

Review the change and the reasons

A first responder for every alert

Every incident improves your observability