Back to product

Technical Details

Live data from a real Syke deployment — 4 sources perceived by Opus 4.6 with the Agent SDK. Anonymized. The agent explores your footprint interactively, not via prompt dump.

3,225events
4sources
94.3%accuracy
$0.60per run
The Process

Watch the agent think

189s of Opus 4.6 exploring a digital footprint with 6 MCP tools

Agent Activity
1/22
0s

Starting perception. I need to understand who this person is from their digital footprint across 3 platforms. Let me begin by surveying what's available.

Source Coverage0%
Claude Code
pending
ChatGPT
pending
GitHub
pending
Tool calls0
Topics found0
Connections0
Elapsed0s / 189s

A perception run, unfolding

55 tool calls. 3 minutes. The agent sweeps all six sources, searches for recurring patterns, cross-references across platforms — then synthesizes a coherent identity model.

Claude WebClaude CodeClaude CoworkGmailChatGPTGitHub
0s / 189s
Coverage0%
Agent Reasoning

Six sources to explore. I'll start with the Claude platforms — they hold the most recent context and usually reveal the builder's current arc.

Coverage gate in action

The agent can't skip platforms. A hooks-based quality gate ensures completeness.

Step 1
Agent submits profile

After exploring the timeline, the agent calls write_section to submit its identity profile.

write_section(section='identity_anchor', content='Alex is a builder...')
Step 2
Coverage gate denies

CoverageTracker detects only 2 of 3 sources explored. Returns PermissionResultDeny.

PermissionResultDeny('Missing sources: github. Explore all platforms before writing.')
Step 3
Agent explores missing source

The agent browses GitHub events, finding 12-year commit history and open-source contributions.

browse_timeline(source='github', limit=50) → 38 events found
Step 4
Profile accepted

With all sources explored, the coverage gate passes. The profile now reflects the complete digital footprint.

write_section(section='identity_anchor') → Coverage: 100% ✓
Learning

The search strategy gets smarter

12 runs, 4 strategy evolutions — ALMA's relevance climbs from rank 4 to rank 1. The v3 dip is exploration cost: new entity searches tested, pruned, replaced.

v1: Concept Search
v2: Topic Expansion
v3: Entity Discovery
v4: Refined Ranking
67% cheaperper run vs cold-start0 LLM callsfor learning3 runsto first strategy
94.3%peak at Run 5 — $0.60, cheaper than the $1.80 baseline
Search Strategy
Concept space stabilized — ALMA reaches #1, cost drops 14%
1ALMA
73%
2acme-labs
73%
3companion
58%
4context-engine
58%
5hackathon
51%
6ContextMem
23%
7memory
16%
8federated
16%
15 queries · $8.07 cumulative
ALMA's Rank
#4
v1
#1
v2
#1
v3
#1
v4
Relevance: 0.596 → 0.732
concept search > entity search

Identity growing over time

The same person, perceived with increasing depth

threads
2
6
6
2p
3p
3p
platforms
Identity Profile — Run 5
3,225 events · claude-opus-4-6
Identity Anchor

Alex is a builder threading the needle between survival pragmatism and civilizational ambition. An engineer who raised $2M in seed funding for Acme Labs (privacy-preserving AI), he arrived in SF in Oct 2025 and is now caught between runway anxiety and audacious technical vision.

Discovered this run
$2M Acme Labs + runway anxiety surfacedALMA meta-learning connection madeChatGPT existential threads cross-referencedSF arrival context grounded
Active Threads
Context Engine: Meta-Learning Perceiver
NEWCompanion App: Autonomous Agent
NEWAI Research Framework at Acme Labs
NEWRunway & City Adaptation
NEWExistential Startup Reflection
NEWAcme Labs Strategic Vision
Voice Patterns

Precise, urgent, self-critical, spiritually curious

The agentic perceiver works — 29 runs today, zero errors, $22 spend, profile updating every 15 min. But every run starts cold.
The dog is not a pet. It is an aspect of the companion's psyche — a mask, a shadow, a fractal of the maker.
ALMA's core insight: the meta-agent evolves executable exploration strategies, not just outputs
Perceived fromchatgptclaude-codegithub