Mergeable by default: Building the context engine to save time and tokens — Peter Werry, Unblocked
Summary
Peter Werry, co-founder of Unblocked, delivers a practitioner's guide to building context engines for AI coding agents — the reasoning layer that supplies the right organizational context to agents so they can produce code that fits an existing system, team conventions, and past decisions. The session is divided into a talk component followed by a live workshop where attendees build a social engineering graph against their own repositories.
Werry opens by situating the problem on an AI adoption curve. Four years ago, developers operated as the context engine themselves, manually supplying agents with issue tickets and code excerpts. Today, most teams are at the "parallel agents + MCP + skills" stage, but the puck is moving toward fully autonomous background agents running in cloud CI systems. The key bottleneck is no longer intelligence — Werry references the Mythic coding model as approaching near-perfect code intelligence — but context. Without rich organizational context, agents enter doom loops: repeated incorrect attempts that waste time and tokens.
The talk systematically dismantles three myths. First, naive RAG over internal docs is not a context engine. Vector search alone triggers what Werry calls the "satisfaction of search" problem — a term borrowed from radiology where doctors stop searching after finding the first plausible diagnosis, missing deeper pathologies. Agents similarly stop retrieving once they find a plausible code reference, missing golden nuggets buried in old Slack threads or incident post-mortems. Second, wiring up a collection of MCP servers does not constitute a context engine; access does not equal understanding. Third, bigger context windows do not solve the problem. Even a million-token window cannot reason across conflicting sources, and most enterprises hold far more than one million tokens of organizational knowledge.
Werry outlines six requirements for a proper context engine: (1) unified system context that builds semantic relationships between data sources — not just hyperlinks but distilled best-practice memories extracted from repeated PR comment patterns; (2) conflict resolution using recency and code-branch biasing, but also forward-looking Slack signals from domain experts; (3) targeted retrieval personalized to a developer's active repositories, using deeper vector retrieval on frequently-contributed repos and broader retrieval elsewhere; (4) data governance that flows access control from the source — private Slack channels are queried only when the requester has permission, and answers are private to that user; (5) expert bottling, where the contributions, PR comments, and Slack discussions of each domain expert are distilled into a reusable context artifact that seeds new agent sessions; and (6) a social engineering graph that surfaces team clusters, expert coverage, and cross-team relationships, built via PageRank-style procedural graph construction layered with ML vector-cluster proximity and LLM-summarized distillations.
A concrete benchmark illustrates the value: implementing Anthropic's new adaptive thinking mode (which replaced the legacy token-budget approach) took 2 hours 25 minutes and consumed 21 million tokens without the context engine. With Unblocked's context engine enabled, the same task took 25 minutes and 10 million tokens — a 6x speed improvement and 52% token reduction, entirely because the agent understood the legacy code it was replacing and avoided doom loops.
Three hard lessons from production: (1) optimizing for data access rather than understanding led to failed retrieval; (2) hiding unresolvable conflicts instead of surfacing them to the user degraded trust and correctness; (3) caching context engine answers and feeding them as context into future queries caused mean-reversion and context poisoning.
The highest-value use cases identified are planning (biggest ROI), code review with organizational motivation, ticket enrichment, production triage wired to Sentry and DataDog, incident management correlating alerts with past Slack discussions, and engineering support channels where the context engine auto-answers questions from other teams. On agent usage share, Claude Code leads by a wide margin among Unblocked customers, followed by Cursor, with Claude Desktop unexpectedly popular and VS Code and Codex accounting for a smaller share.
Key Takeaways
- •Access to data sources does not equal understanding — wiring up MCP servers without a reasoning layer leaves agents unable to grasp why decisions were made or what failed previously.
- •The satisfaction of search problem causes agents to stop retrieving once they find a plausible result; a context engine must push past that first hit to surface Slack threads, incident reports, and deleted code history.
- •Bigger context windows are not a substitute: most enterprises hold far more than one million tokens of organizational knowledge, and even large windows cannot resolve conflicting sources or select the right information without additional reasoning.
- •Pre-seed agents with expert context at the start of a session — bottled expert profiles (PRs, Slack signals, code ownership) act as a jump-off point that directs the rest of retrieval and cuts implementation time dramatically.
- •Use recency and code-branch biasing as a starting point for conflict resolution, but bias toward forward-looking expert Slack conversations when the goal is where the system is going, not where it has been.
- •Never cache context engine answers to serve future similar queries — code and decisions change constantly, and reusing previous answers causes mean-reversion and context poisoning.
- •Surface unresolvable conflicts to the user rather than silently resolving them; human feedback on these signals is high-volume and valuable for continuously improving the engine.
- •Personalize retrieval by weighting vector search toward repositories where the requesting developer has the most PR contributions, using deep retrieval on focused repos and shallow retrieval on the rest.
- •Flow access control from the data source up through the context engine — private Slack channels and restricted repos should only contribute to answers for users who have permission, and those answers must not be shared publicly.
Transcript
All right, thanks everyone. Sorry about the wait. Um This is going to be a bit of a strange session because there is a workshop component to this. So I guess everyone will be coding on their laps. Sorry. Um but anyway, sorry, I'm here and this is my colleague Brandon. Um so we're we're going to break this this session into two different parts. One is a sort of a talk that I'm going to give about what context engines are are useful for and and how you might go about building one, what to think ab...
Speakers
Topics
Related Sessions
