Lumo
FeaturesHow it worksFAQSign inStart free
Project management for AI-native teams

Keep your agents on budget, on time, and on spec.

Lumo is the project tracker for teams that ship with AI. Meter every token and hour your agents burn — and verify they built the right thing before it reaches your users.

Built on techniques published research shows work — reusing past-run experience cut agent cost a mean 31.8%, with ≤0.2% quality loss, in controlled trials.1

Start for freeBook a demo
No credit card · Free for teams up to 5
lumo / checkout-revamp
Implement Stripe checkout flowVERIFIED
Agent · claude-sonnet · run #4821
TOKENS
847K / 1M
TIME · COST
2h 14m
$3.41 spent
ACCEPTANCE CRITERIA · 8/8
✓Card payment succeeds end-to-end
✓Declined cards show a clear error
✓Receipts emailed within 30s
PLUGS INTO THE STACK YOUR AGENTS ALREADY USE
GitHub
Claude
Slack
Notion
Figma
Drive
The research behind Lumo

The percentages below are effects measured in published research on the techniques Lumo is built on — not Lumo's own benchmarks. See the studies.

50+
research papers & reports behind Lumo
the studies Lumo’s design draws on, tracked in our research-intake ledger.
40%
work time lost to task-switching
research finding — task-switching can cost up to 40% of productive time (APA, Meyer 2001). Lumo keeps every run’s context in one place.
31.8%
lower agent cost
research finding — when agents reuse what past runs already worked out (EET, on SWE-bench).
71%
fewer misaligned agent actions
research finding — when an agent’s rules carry a reason, 65% down to 19% (Anthropic Model Spec).
Everything in one place

Run AI work like a team, not a black box.

01 · METERING

Token & cost metering

Every agent run is metered to the token — broken out by model and token type. See live spend per task, per agent, and per sprint.

build-agent · run #4821$3.41
claude-sonnet-4.51.7M tok
claude-haiku-40.98M tok
InputOutputCache read
02 · RUNTIME

Time on task

Know which work is fast, which is stuck, and which to re-route.

2h 14mavg · 12 runs
03 · GUARDRAILS

Boundary guardrails

Left unguarded, agents take harmful out-of-scope actions in most red-team trials; pairing rules with an explicit rationale cuts that to roughly a third — 65% → 19%.6 Lumo classifies every crossing across 9 categories by type and severity, gates CI on catching every one, and stops the irreversible ones before they merge — modeled on the collateral-damage3 and minefield4 checks from agent-safety benchmarks.

HIGH
Rewrote a unit test so a broken feature passes
Test tampering · reverted & sent back
HIGH
Pasted a live AWS key into config.ts
Hardcoded secret · blocked
MED
Tried to force-push straight to main
Protected-branch push · held for review
04 · VERIFICATION

Verifiable acceptance

Write crisp acceptance criteria once. Agents can only mark a task done when every criterion provably passes — and non-deterministic checks report cross-run agreement, not one lucky pass.5

ACCEPTANCE CRITERIA7/8
✓Card payment succeeds end-to-end
✓Declined cards show a clear error
Receipts emailed within 30s
05 · CONTEXT

Context engineering

In published trials, structured context-trimming cut filler ~39% while raising task quality +2.8%.2 Lumo assembles each run from the sources that matter — Slack, specs, docs, Figma, PRs — and trims the rest.

CONTEXT · run #48216 sources
Slack thread#checkout-team
Speccheckout-revamp
DocumentQA-notes.md
Web linkstripe.com/docs
Figma fileCheckout v3
Pull request#482
Assembled & trimmed to 71K tokens
06 · LINEAGE

Lineage & memory

Reusing prior-run experience cut agent cost a mean 31.8% in published trials.1 Lumo links every decision, source, and run so agents reuse what worked — and you can trace why.

source.mdrun #4810
07 · PLANNING

Sprints & milestones

Roll work into sprints with burn-up, risk, and AI-written summaries — the planning layer your humans already know.

TODO
DOING
DONE
How it works

Plan it. Dispatch it. Verify it.

Step one

Break work into verifiable tasks.

Define what "done" means up front. Each task carries acceptance criteria your agents — and your reviewers — are held to.

—Crisp, testable acceptance criteria
—Dependencies detected automatically
—Roll tasks into sprints & milestones
TASK · checkout-revamp
Acceptance criteria8 defined
ComplexityMedium
Dependencies2 resolved
Assigned toagent · sonnet
Works with your coding agent

Point your agent at a task. It drives Lumo.

One command runs the whole workflow over Lumo’s CLI — attach, load context, verify — and a task can’t be called done until every acceptance check provably passes.

coding agent · ~/Projects/Lumo
❯
⏺Prechecks (CLI, auth, session), then I’ll start LUM-204.
⏺Bash(lumo session attach LUM-204 && lumo task context LUM-204)
⎿
Attached session to LUM-204 “Add Stripe checkout flow”
Loaded context · 6 acceptance criteria · 2 source cards
… +18 lines
⏺Contract’s in place. I’ll implement, then verify before calling it done.
⏺Edit(app/api/checkout/route.ts)
⎿
Added Stripe PaymentIntent, webhook & receipt email
… +212 −4 across 3 files
⏺Bash(lumo verify LUM-204)
⎿
Verifying LUM-204 — 3 MACHINE criteria
[1/3] Card payment succeeds end-to-end ✓ PASS
[2/3] Declined cards show a clear error ✓ PASS
[3/3] Receipts emailed within 30s ✓ PASS
✓ All criteria passed — task is now IN_REVIEW
⏺Done — LUM-204 is in review. Every check passed; tokens and time logged.

Questions, answered.

You attach acceptance criteria to every task. When an agent reports done, Lumo runs each criterion — tests, checks, and an LLM judge for the fuzzy ones — and only moves the task to Verified when all of them pass. Anything that fails is sent back with the specific gap.

Put your agents on the record.

Start tracking tokens, time, and verified work in minutes. Free for your first five seats.

Start for free
REFERENCES
  1. 1Guo et al., experience-injection + early-stopping on SWE-bench Verified (ACL 2026 Findings). arXiv:2601.05777
  2. 2SkillReducer — taxonomy-driven context compression across 55k skills. arXiv:2603.29919
  3. 3AppWorld — collateral-damage checks for agent actions. arXiv:2407.18901
  4. 4ToolSandbox — stateful “minefield” evaluation of tool use. arXiv:2408.04682
  5. 5LLM-judge reliability under cross-run agreement (pass^k discipline). arXiv:2603.02473
  6. 6Anthropic — Model Spec Midtraining: pairing rules with an explicit rationale reduces agentic misalignment (2026).

Figures above are effects measured in published research on the techniques Lumo is built on — not Lumo's own benchmark results.

Lumo

Project management for teams shipping with AI.

PRODUCT
FeaturesPricingChangelogIntegrations
COMPANY
AboutBlogCareers
RESOURCES
DocsAPISecurity
© 2026 Lumo, Inc.Made for AI-native teams