Nathan Gallo

Selected work

• Production AI

  Microsoft Teams AI assistant

  An assistant inside Microsoft Teams, built with Azure OpenAI, Semantic Kernel and .NET, that prepares customer briefs.

  Before

  3+ hours

  After

  15 minutes or less

  Reported times for this one workflow, not a controlled study. The code and data belong to the client and aren't published.

  Case study

• Developer tooling

  Codebase Memory CLI

  A command-line tool that indexes a repository so people and coding agents can find code.

  A fork of DeusData/codebase-memory-mcp. The indexing engine is DeusData's work; the command-line version is mine.

  $ time codebase-memory-cli index . \
      --name agent-workflow --mode fast \
      | jq -c '{project, nodes, edges, status}'
  {"project":"agent-workflow","nodes":6938,"edges":20726,"status":"indexed"}
  real    0m4.809s
  $ codebase-memory-cli trace \
      --project agent-workflow \
      --function-name _accepted_run \
      --direction inbound \
      | jq -c '{callers_total}, (.callers.groups[] | {qn_prefix, rows})'
  {"callers_total":4}
  {"qn_prefix":"agent-workflow.src.agent_workflow.task_integration","rows":[["integrate",1]]}
  {"qn_prefix":"agent-workflow.src.agent_workflow.task_chains","rows":[["operate",1]]}
  …
  

  Case studySource ↗

• Agent evaluation

  Agent-Workflow and its evaluation

  Agent-Workflow runs coding agents and keeps a sealed record of each run.

  Successful runRecovery

  Codex fixes a one-line bug. Agent-Workflow runs the tests itself before it records the run as complete.

  Prepare. Fixes the task, the source commit and any acceptance commands before the worker starts.

  ‹Replay›1 / 9

  1. +0:00prepared: Agent Run initialized; source 7ff9ef4Prepared the run: the task, the source commit and the acceptance command are fixed before the worker starts.
  2. +0:00role implementation; class implementationI named the role when preparing this run, so the router was not consulted. The role sets the worker's model and permissions.
  3. +0:01prepared -> running: headless worker startedAgent-Workflow started Codex as a child process it owns.
  4. ·exit 0: Ran 3 tests in 0.000s OKThe worker fixed the bug and ran the tests itself.
  5. ·[main 629ce21] Fix minute duration parsingThe worker committed the fix.
  6. ·state: finalized | result: completedAgent-Workflow ran the declared acceptance command itself before accepting the worker's completion.
  7. ·acceptance:unit-tests: passThe pass comes from that command's exit code and output hashes, not from the worker's report.
  8. +0:34completed; 38 artifacts hashedSealed: every evidence file is hashed into a read-only receipt.
  9. ·disposition: NoneCompleted is not accepted. The run waits here until a person reviews it.

  Run duration-fix-cleanSource runs/duration-fix-clean/events.jsonl · sequence 1

  Recorded output ↗

  

  Close

  Recorded on 6 November 2026 (UTC) with Agent-Workflow 0.13.0 (commit a7fc48f) and codex-cli 0.166.0, routing in deterministic mode. Monospace lines are copied from Agent-Workflow's output; the plain text under each is my note. A · means the worker's log records order but not time.

  Identified the kind of task

  Model-based82%

  Fixed rules53%

  Spotted when a person was needed

  Model-based92%

  Fixed rules72%

  120 test cases. On rating risk, the two were about even. It doesn't show the model is better in general.

  Agent-WorkflowEvaluationResults ↗