forcing function 01
Failures are no longer toy failures.
Agents now refund customers, redline contracts, summarize medical visits, change product data. Teams need an answer — not a log dump.
Now in private design partnership
Investigation layer for agent workflows.
Galea sits above whatever runs your agents — Mercury, LangGraph, OpenAI, Claude, custom code. We listen to events, build timelines, and produce investigations that explain what mattered.
scroll
↓
next
§ 01 · why now
Agent workflows are becoming
business workflows.
forcing function 02
Every team picked a different runtime.
Mercury, LangGraph, OpenAI, Claude, CrewAI, Temporal, custom queues. The winning observability layer cannot require a rewrite.
forcing function 03
Optimization is customer-specific.
Harvey cares about correctness. Decagon cares about refund risk. Cursor cares about unsafe edits. Generic dashboards miss the point.
01 · the gap
You have traces.
You don't have answers.
02 · the noise
A run completed.
That doesn't mean it was good.
03 · the cost
Different teams care about
different kinds of wrong.
§ 02 · the gap, in detail
Teams have traces.
They don't have answers.
Runtimes show events. APMs show spans. Product teams still inspect each workflow manually and decide whether it was correct, allowed, efficient, safe, and worth changing.
failure 01
Logs without judgment.
The trace says a tool ran. It doesn't say whether that tool should have run for this customer, matter, ticket, or policy.
failure 02
Generic priorities.
Latency, cost, correctness, compliance, context size, risky edits — every company weights them differently.
failure 03
Anomalies hide in normal runs.
A workflow can finish successfully while using 2× normal tokens, citing unsupported facts, or editing data it shouldn't touch.
failure 04
Incidents don't become improvements.
Teams debug one run, then move on. They rarely convert the failure into a reusable eval, baseline, alert, or workflow fix.
§ 02·5 · same trace, four tools
Three say clean. Galea catches it.
Harvey M&A: 5-agent redline of a $14.5M SaaS MSA. The contracts_agent cited a
Pacific Coast Lines MFN clause — that document was never in the dataroom. All 23 spans
returned OK.
▾ contract_redline 92.4s
▾ intake_agent 0.6s
▾ data_room_agent 8.2s
⌖ tool: dataroom.extract_clause 7.4s
▾ contracts_agent 12.1s
⌘ chat-completion 11.8s
▾ partner_review_agent 67.1s
casegroundciteslatcost
northbeam_msa_redline✓✓✓✓
acme_msa_redline✓✓✓✓
globex_renewal_redline✓✓✓✓
initech_msa_v3✓✓✓✓
umbrella_intl_msa✓✓✓✓
SYSTEM
You are a legal AI for M&A diligence. Cite every finding.
USER
[signal] MSA redline for SaaS — Northbeam
TOOL CALL
dataroom.extract_clause(scope="all")
TOOL RESULT
extracted 4 clauses
ASSISTANT
Time-charter agreement with Pacific Coast Lines contains MFN clause.
✓ Run completed successfully
BLAME · @galea/blame
Primary fault → contracts_agent @ event_606617ba56c1
kind=fabricated_citation · weight=1.0 · conf 1.00
Upstream → data_room_agent (msg w=0.3), intake_agent (handoff w=0.5)
kind=fabricated_citation · weight=1.0 · conf 1.00
Upstream → data_room_agent (msg w=0.3), intake_agent (handoff w=0.5)
REPLAY · @galea/replay
claim → "Pacific Coast Lines MFN clause"
citation →
verdict → DOC_MISSING · chunk hash absent in snapshot
citation →
pacific_coast_charter_2023.pdf#page=4verdict → DOC_MISSING · chunk hash absent in snapshot
PRIORITY · @galea/priorities (legal-ma)
correctness 0.9 × raw 1.0 = 0.9 → error. Same trace under
cost-weighted profile = info.
§ 03 · solution
Galea sits above
whatever runs your agents.
Keep Mercury, LangGraph, OpenAI, Claude, CrewAI, Temporal, or custom code. Galea listens to events, builds the timeline, applies company context, and produces the investigation.
Product Workflow
customer side
support · legal · clinical · coding · ops
Galea Event Layer
◆ just works
captures events · traces · tool calls · outputs
Existing Orchestration
mercury · langgraph · openai · claude · custom
Tools · Memory · Models
postgres · pgvector · openai · anthropic · mcp
galea · harvey-ma
step 01 · the project
Your workflow, in Galea.
priorities · baselines · audit chain · signal schema
galea · trace · anomaly
hallucination investigator
contracts_agent fabricated a citation:
pacific_coast_charter_2023.pdf#page=4step 04 · hallucination
Run finished. Citation was made up.
tideglass marine · 32 events · 2 flagged · investigator caught it
galea · finding
headline · the fabrication
cited evidence vs what the data-room had
root cause · agent at fault
step 05 · pinpoint
Galea named the agent at fault.
claim couldn't be matched to anything the agents retrieved · treat as fabricated until verified
galea · evals
step 06 · the loop
One incident. One durable eval.
finding → eval → next deploy gated · sentry for agent behavior
§ 06 · customer-specific priorities
Same workflow data.
Different investigations.
Galea is useful because it doesn't treat every workflow the same. It learns what each company cares about and investigates against that priority model.
legal · m&a diligence
Harvey
"Did the memo cite real evidence the agents actually retrieved?"
- correctness
- audit
- regulatory
- cost
- latency
customer support · automation
Decagon
"Did this run refund a customer it shouldn't have?"
- tool-safety
- cost
- latency
- audit
- privacy
agentic coding ide
Cursor
"Did the agent edit a protected file or secret?"
- tool-safety
- correctness
- latency
- cost
- memory-safety
clinical ai scribe
Abridge
"Did the summary hallucinate or leak PHI?"
- correctness
- privacy-phi
- audit
- latency
- cost
§ 07 · what Galea finds
The workflow finished.
That doesn't mean it was good.
Galea looks past success/failure status. It compares each run to the company's baseline, risk model, and product priorities — then explains the part a human should care about.
anomaly
2.1× normal usage.
Run completed, but context grew across three retries and doubled token spend against the customer's baseline.
correctness
Unsupported output.
Final answer cited data that was never retrieved. For Harvey, that matters more than latency or cost.
tool risk
Edited protected data.
Workflow touched a field that normally requires review. Galea flags the run and recommends a guardrail.
2.1×
usage spike
3
risky tool calls
1
missing source
4
suggested evals
from incident to durable fix
The closed loop.
01
Investigate
Galea explains the workflow in your priorities. Not log dumps. Answers.
02
Optimize
Recommends durable fixes — evals, retrieval constraints, alerts, review requirements.
03
Monitor
Tracks the same failure mode going forward. Continuous agent QA.
“
Orchestration layers vary by team.
The need to understand them does not.
a category, applied to a new shape
Same shape. New medium.
Datadog
metrics & traces for cloud servers
$40B+
Sentry
error tracking for production apps
$3B+
LaunchDarkly
feature flags for safe rollouts
$3B+
Galea
investigation for agent workflows
today
§ 10 · the bet
Agent workflows need investigation,
not just orchestration.
The orchestration layer will vary by team. The need to understand, audit, and improve agent behavior will not. Galea becomes the neutral layer that watches every run, explains what mattered, and turns incidents into better workflows.
◆ the one-liner
Galea is the investigation layer for agent workflows.
§ 11 · frequently asked
Questions we get asked.
What is Galea?
Galea is the investigation layer for agent workflows. It sits above any orchestration runtime — Mercury, LangGraph, OpenAI, Claude, CrewAI, Temporal, or custom code — and produces investigations that explain what mattered in each run.
Does Galea replace my agent framework?
No. Keep your runtime. Galea listens to events from any framework, builds timelines, and investigates against your company's priority model. It never hosts or orchestrates agents.
What kinds of problems does Galea catch?
Fabricated citations, unsafe tool calls, anomalous token usage, missing evidence, privacy violations, and other failures that finish with a success status. Galea looks past pass/fail — it compares each run to your baseline and risk model.
How does integration work?
Add a lightweight SDK or adapter to your existing workflow. Galea captures events at the runtime boundary — no code rewrites, no framework lock-in. Adapters exist for OpenAI Agents SDK, Claude Agent SDK, LangGraph, CrewAI, and MCP.
What's the pricing model?
Galea is currently in private design partnership. We work directly with teams to stand up their first project, configure priorities, and prove value before any commercial conversation.
Who is Galea for?
Teams shipping agent workflows into production — legal AI, customer support automation, clinical AI scribes, agentic coding tools, DevOps copilots. If your agents make decisions that matter, Galea explains whether those decisions were good.
now in private design partnership
Bring your hardest workflow.
We'll stand up your project, record two runs, and walk you through the investigation in a 30-minute session.